Prediction of cognitive impairment through speech data analysis: A comparative evaluation of deep learning models

BACKGROUND: The early detection of cognitive impairments, such as mild cognitive impairment (MCI) and Alzheimer's disease (AD), is essential for timely intervention and management. This study evaluates the performance of various deep-learning models in classifying speech recordings from individuals with normal cognition (NC), MCI, and AD, to identify the most effective approach for audio-based cognitive impairment diagnosis. METHODS: Speech data were obtained from the AI Hub "Cognitive Impairment Diagnosis Voice/Conversation" dataset. The study analyzed voice recordings from 320 female participants (105 with Alzheimer's disease, 92 with mild cognitive impairment, and 123 cognitively normal controls). Three deep-learning architectures were compared: a one-dimensional convolutional neural network (1D CNN), an audio spectrogram transformer (AST), and a speech recognition model (Wav2Vec 2.0). The models were trained using features such as spectrograms, mel-spectrograms, and mel-frequency cepstral coefficients (MFCCs). Model performance was assessed using accuracy, recall, precision, and F1-score, with a five-fold cross-validation strategy to ensure robust and unbiased evaluation. Statistical significance was assessed using pairwise proportion z-tests with Holm-Bonferroni correction, and Wilson score 95% confidence intervals were computed for each model's accuracy. RESULTS: Wav2Vec 2.0 outperformed the other models, achieving the highest accuracy and F1-scores for NC vs. MCI (accuracy: 0.74, F1: 0.72) and NC vs. AD (accuracy: 0.83, F1: 0.83). Pairwise proportion z-tests with Holm-Bonferroni correction confirmed that Wav2Vec 2.0 significantly outperformed 7 of 10 competing models (corrected p < 0.05) in both classification tasks. Performance varied by model and classification task, with Wav2Vec 2.0 consistently demonstrating superior accuracy across labels. CONCLUSION: This study emphasizes the importance of selecting appropriate models and features for task-specific optimization and provides a foundation for developing non-invasive, speech-based diagnostic tools for cognitive disorders.