Biomarkers.

Area of Science:

• Neurology
• Computational Linguistics
• Biomedical Engineering

Background:

• Alzheimer's disease (AD) and dementia with Lewy body (DLB) share early symptoms like memory and cognitive issues.
• Differentiating AD from DLB is challenging for large-scale patient screening.
• Automated classification using speech features is proposed to distinguish AD from DLB.

Purpose of the Study:

• To develop and evaluate automatic classification systems for distinguishing Alzheimer's disease (AD) from dementia with Lewy body (DLB).
• To investigate the efficacy of combining text and acoustic speech features for differential diagnosis.
• To facilitate scalable screening of patients for clinical trials and interventions.

Main Methods:

• Collected speech data from patients with AD (ADCI) and DLB (DLBCI), and healthy controls (HC).
• Extracted text embeddings using a large language model (KLUE-RoBERTa-base) and acoustic features using OpenSMILE.
• Applied Principal Component Analysis for feature reduction and experimented with data-level and decision-level fusion for Multilayer Perceptron classifiers.

Main Results:

• The fused model (text + acoustic features) achieved 75.7% accuracy (AUC=0.77) for distinguishing HC from all patients.
• For ADCI vs. DLBCI classification, decision-level fusion yielded the highest performance with 80.3% accuracy (AUC=0.82).
• Fused models outperformed text-only or acoustic-only approaches.

Conclusions:

• Automatic classification of ADCI and DLBCI using speech analysis is a promising tool.
• This method can significantly aid in the large-scale screening of patients for clinical trials.
• Speech-based analysis offers a viable non-invasive approach for differential diagnosis.