Enhancing dementia and cognitive decline detection with large language models and speech representation learning.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Signal Processing

Background:

• Dementia presents a significant global health challenge.
• Early detection of cognitive decline is crucial for timely intervention.
• Spontaneous speech analysis offers a non-invasive diagnostic approach.

Purpose of the Study:

• To predict cognitive decline from speech samples for the PROCESS Signal Processing Grand Challenge (ICASSP 2025).
• To evaluate the efficacy of integrating advanced AI models with speech features for cognitive assessment.

Main Methods:

• Utilized eGeMAPS features from openSMILE, HuBERT (self-supervised speech representation), and GPT-4o (large language model).
• Integrated features with custom LSTM and ResMLP neural networks and Scikit-learn regressors/classifiers.
• Trained models on speech data from 157 participants and evaluated on a separate test set of 40 individuals.

Main Results:

• The LightGBM regression model achieved an RMSE of 2.7775, ranking 10th globally.
• The LSTM classification model obtained an F1-score of 0.5521 for dementia/MCI/control, ranking 20th globally.
• The proposed method surpassed baseline models in predicting cognitive decline.

Conclusions:

• Integrating large language models with self-supervised speech representations enhances cognitive decline detection.
• The approach provides a scalable, data-driven method for early cognitive screening.
• This technology may support future applications in neuropsychological informatics.