Predicting Memory Score Using Paralinguistic Features.

Area of Science:

• Neurology
• Computational Linguistics
• Gerontology

Background:

• Poor short-term memory is an early indicator of dementia risk.
• Existing speech analysis methods struggle to differentiate early dementia from normal aging and predict dementia risk.
• The LOGOS episodic memory test is a feasible platform for assessing memory function over the phone.

Purpose of the Study:

• To develop and validate an automated method using speech paralinguistic features to predict dementia risk.
• To assess the ability of speech features to discriminate between individuals with strong and poor short-term memory performance.
• To evaluate the efficacy of this method across multiple datasets and feature selection techniques.

Main Methods:

• Extraction of paralinguistic features from audio recordings of individuals undergoing the LOGOS episodic memory test.
• Application of various feature selection methods to identify the most discriminative speech characteristics.
• Utilizing a Support Vector Machine (SVM) classifier to predict short-term memory performance based on extracted features.

Main Results:

• The best performing model, an SVM classifier, achieved an accuracy of 84% in predicting short-term memory performance per audio recording.
• Paralinguistic speech features effectively discriminated between individuals with strong and poor short-term memory.
• The method demonstrated robustness across multiple datasets.

Conclusions:

• Automated analysis of speech during the LOGOS episodic memory test can effectively estimate dementia risk.
• This non-invasive method shows clinical relevance for early dementia risk assessment in older adults.
• Speech-based dementia risk prediction offers a promising avenue for early detection and intervention.