Biomarkers.

Area of Science:

• Neurology
• Computational Linguistics
• Gerontology

Background:

• Early detection of mild cognitive impairment due to Alzheimer's disease (MCI d/t AD) and AD dementia (ADD) is crucial for timely intervention.
• Speech analysis offers a non-invasive method to identify subtle cognitive deficits.
• This study investigates acoustic and lexical speech features for classifying cognitive status in older Korean adults.

Purpose of the Study:

• To evaluate the utility of speech analysis in classifying older Korean adults across three clinical scenarios: screening for AD, detecting advanced ADD, and differentiating healthy controls (HC) from ADD.
• To demonstrate the feasibility of speech-based methods for supporting earlier and more targeted interventions for cognitive impairment.

Main Methods:

• 110 older Korean adults (HC, MCI d/t AD, ADD) were recruited. Speech data were collected from neuropsychological tests and specific tasks.
• Acoustic features (openSMILE) and lexical features (Korean RoBERTa) were extracted and processed using Principal Component Analysis.
• Three multilayer perceptron models (acoustic-only, lexical-only, ensemble) were trained and evaluated using 5-fold cross-validation.

Main Results:

• Ensemble models combining acoustic and lexical features outperformed single-feature models across all classification tasks.
• Accuracy for HC vs. AD was 75.8% (AUC 0.756), Non-dementia vs. ADD was 85.1% (AUC 0.801), and HC vs. ADD was 87.0% (AUC 0.893).
• The combination of features provided complementary information, reflecting both vocal characteristics and language-based deficits.

Conclusions:

• Speech-derived features effectively detect cognitive impairment in older Korean adults across various diagnostic scenarios.
• This non-invasive approach can facilitate earlier, more targeted interventions.
• The method shows potential to simplify clinical workflows and expand screening accessibility, especially in resource-limited settings.