Biomarkers.

Area of Science:

• Computational linguistics
• Biomedical engineering
• Neuroscience

Background:

• Alzheimer's disease (AD) diagnosis relies on invasive and costly methods, limiting early detection.
• Speech analysis offers a non-invasive alternative for cognitive impairment detection.
• Chinese homophones pose transcription challenges, potentially reducing speech-based model accuracy.

Purpose of the Study:

• To develop and validate a novel speech-based method for improved Alzheimer's disease detection in Chinese speakers.
• To overcome transcription ambiguities caused by homophones by utilizing Pinyin sequences.
• To enhance the efficiency and accuracy of early cognitive impairment screening.

Main Methods:

• Utilized audio data from 300 participants in the China Preclinical Alzheimer's Disease Study (C-PAS) cohort.
• Transcribed speech to Pinyin sequences and extracted acoustic features (e.g., pause frequency, MFCCs).
• Employed a Bi-directional LSTM model to integrate Pinyin sequences, acoustic features, and demographic data, incorporating data augmentation techniques.

Main Results:

• The proposed method achieved 93.80% accuracy and an AUC of 0.93.
• Integrating Pinyin sequences with acoustic features significantly outperformed models using only acoustic data or cognitive scores.
• Ablation studies confirmed the importance of combining linguistic (Pinyin) and acoustic features.

Conclusions:

• Integrating Pinyin sequences and acoustic features is a feasible and effective strategy for non-invasive Alzheimer's detection in Chinese.
• This approach provides a practical tool for early Alzheimer's screening.
• Findings support further large-scale studies and potential clinical applications for cognitive impairment detection.