Alzheimer’s detection using audio through deep learning techniques

Alzheimer’s Disease (AD) is a long-lasting neurodegenerative disorder that progressively weakens memory, communication, and cognitive abilities. Early detection and diagnosis are critical for timely intervention. Recently, Artificial Intelligence (AI) models with speech analysis have emerged in detecting AD using acoustic and prosodic characteristics of individuals. However, they still achieve lower accuracy due to a lack of temporal dynamics and complex nonlinear correlation features among the speech signals. Hence, a new robust, noninvasive, and cost-effective model is proposed in this article for early detection of AD from speech signals. A new hybrid model called TempoBoostNet is developed to simultaneously capture both temporal dynamics and complex nonlinear relationships among speech signals. First, a dataset of 547 speech recordings (247 CE and 300 Healthy Controls (HC)) is gathered from publicly available Kaggle and GitHub sources. The raw audio files are preprocessed using the Wiener filter and Mahalanobis distance to eliminate noise and silence, respectively. Then, various features that reflect both low-level acoustic characteristics and higher-level speech dynamics are extracted. Also, contextual embeddings from Wav2Vec2.0 and Perceptual Linear Prediction (PLP) coefficients are extracted. All extracted features are concatenated to form a unified multi-level feature set for better feature representation. Moreover, this feature set is utilized to train the TempoBoostNet, which is built by hybridizing the Bidirectional Long Short-Term Memory and Extreme Gradient Boosting (BiLSTM-XGBoost) classifier for AD detection. Finally, experimental results show that this TempoBoostNet achieves 97.4% accuracy, 97.4% precision, 97.3% recall, and 97.3% F1-score, outperforming traditional models.