Towards Computer-Based Automated Screening of Dementia Through Spontaneous Speech

Area of Science:

• Neurology
• Computer Science
• Artificial Intelligence

Background:

• Dementia, particularly Alzheimer's disease, significantly impacts individuals and society.
• Early detection of dementia is crucial for timely intervention and management.
• Spontaneous speech analysis presents a promising avenue for non-invasive dementia screening.

Purpose of the Study:

• To evaluate the effectiveness of machine learning models in classifying Alzheimer's dementia using spontaneous speech data.
• To compare the performance of a pretrained audio feature extractor (VGGish) with custom-built models.
• To introduce and assess a novel convolutional neural network (DemCNN) for dementia detection from raw speech waveforms.

Main Methods:

• Utilized the Alzheimer's Dementia Recognition through Spontaneous Speech (ADReSS) Challenge dataset.
• Employed VGGish, a deep pretrained model, for audio feature extraction, followed by Scikit-learn classifiers (LinearSVM, Perceptron, 1NN).
• Developed DemCNN, a PyTorch-based convolutional neural network operating directly on raw audio waveforms.

Main Results:

• VGGish with Scikit-learn classifiers achieved 59.1% accuracy, outperforming baseline acoustic feature models by 3%.
• The proposed DemCNN model reached 63.6% accuracy, surpassing the best baseline linear discriminant analysis model by 7%.
• Audio transfer learning with VGGish demonstrated superior performance compared to traditional acoustic feature extraction methods.

Conclusions:

• Machine learning, particularly audio transfer learning and deep convolutional neural networks, shows significant potential for computer-based dementia screening.
• The developed methods offer innovative and more effective approaches for detecting dementia through spontaneous speech analysis.
• These findings contribute to advancing non-invasive diagnostic tools for early Alzheimer's dementia detection.