Biomarkers.

Area of Science:

• Artificial Intelligence
• Computational Linguistics
• Neuroscience

Background:

• Accessible and scalable diagnostic methods for cognitive impairment and Alzheimer's disease are crucial.
• Deep learning approaches analyzing speech offer a promising avenue for early detection.
• Current diagnostic tools can be invasive or costly, highlighting the need for non-invasive methods.

Purpose of the Study:

• To develop and evaluate a deep learning model for early detection of cognitive decline using speech analysis.
• To compare the efficacy of acoustic versus linguistic speech features in predicting cognitive impairment.
• To assess the performance of a Bidirectional Long Short-term Memory (BiLSTM) model in classifying mild cognitive impairment (MCI) and unimpaired cognition (UC).

Main Methods:

• Speech samples from 81 participants (19 MCI, 62 UC) were collected from the Alzheimer's Disease Research Center (ADRC).
• A Bidirectional Long Short-term Memory (BiLSTM) model was trained on acoustic features (MFCCs, Spectral Centroid, Spectral Contrast) and linguistic features (vocabulary richness, sentiment, etc.).
• The model utilized 80% of data for training, 10% for validation, and 10% for testing, with 2-fold cross-validation to prevent overfitting.

Main Results:

• The BiLSTM model achieved 87.50% accuracy and a 93.33% F1 score using acoustic features.
• Using linguistic features, the model attained 81.25% accuracy and an 88.89% F1 score.
• The proposed BiLSTM model demonstrated superior performance compared to traditional machine learning and standard LSTM models.

Conclusions:

• Acoustic speech features are more effective predictors of cognitive impairment than linguistic features.
• The developed BiLSTM model shows significant potential for improving the clinical diagnosis of cognitive impairment.
• The approach offers an easy-to-implement and scalable solution for early cognitive decline detection.