An effective and deep learning model to predict Alzheimer’s disease at an early stage from audio data

A leading cause of dementia, called Alzheimer’s disease (AD), poses a challenge to both patients and their families in multiple ways. Thus, the early identification of AD using symptoms is essential to provide effective treatment and medication. Traditional methods, including scale tests, electroencephalograms, magnetic resonance imaging, and audio analysis, offer a scalable approach for the automatic detection of AD and are gaining popularity among researchers. DL-based approaches are mainly applied to audio analysis for AD and have reached better results. This study proposes an optimized deep learning (DL) model for AD prediction from audio data by employing an effective embedding mechanism. Initially, the collected audio data from the DementiaBank Pitt Corpus dataset undergoes preprocessing to enhance the dataset’s quality, which includes noise filtering using the Weiner filter and then data augmentation to strengthen the dataset’s diversity. After that, the Dilated Convoluted Wav2vec (DCWav2vec) extracts useful pre-processed data features. Finally, the AD classification is performed using the Lyrebird Optimized Gated Recurrent Unit (LOGRU) method, in which the hyperparameters are optimally selected using the Latin hypercube with non-convergence factor centered Lyrebird optimization (LNLBO) algorithm. The experimental results showed that our model can detect AD with an average accuracy of 97.59% and a specificity of 97.57%, which is higher than that of existing models. Finally, this research study concludes that audio-based methods have the potential to be a promising tool for the early detection of dementia.