Predicting dementia through audio: Ensemble and deep learning approaches using acoustic features.

Area of Science:

• Neurology
• Artificial Intelligence
• Speech Science

Background:

• Dementia is a cognitive decline beyond normal aging, affecting memory, reasoning, and daily activities.
• Communication difficulties are a significant challenge for elderly individuals with dementia.
• Early diagnosis of dementia is crucial for timely intervention and management.

Purpose of the Study:

• To investigate the use of audio recordings and machine learning for early dementia diagnosis.
• To identify key acoustic features indicative of dementia.
• To compare the performance of ensemble learning and deep learning models in dementia detection.

Main Methods:

• Extraction of acoustic features (pitch, loudness, spectral centroid, MFCC, F0) from patient audio recordings.
• Training and evaluation of ensemble models (Random Forest, AdaBoost, XGBoost, Gradient Boost) and deep learning models (BiLSTM, LSTM, CNN-LSTM).
• Hyperparameter tuning, regularization, and cross-validation were employed to optimize model performance.

Main Results:

• The Gradient Boost model achieved the highest accuracy of 90.5% in diagnosing dementia using spectral centroid, MFCC, and F0 features.
• Ensemble learning models demonstrated superior performance over deep learning models in this specific application.
• The study identified specific acoustic patterns associated with dementia in speech.

Conclusions:

• Audio analysis combined with machine learning offers a promising avenue for early dementia detection.
• Gradient Boost and other ensemble methods are effective for diagnosing dementia from speech characteristics.
• Further research is warranted to understand why ensemble models outperform deep learning in certain scenarios.