Predicting Alzheimer’s Disease via Linguistic Patterns in Speech Transcripts

Abstract

Research in artificial intelligence portray that speech and language features with subtle alterations can be used as early, non-invasive markers of cognitive impairment, yielding a promising second choice for scalable and low-cost screening of Alzheimer's disease. In this paper, we propose an alternative approach that is cost-effective and try to classify Alzheimer versus healthy subjects based on their speech-based text responses analyzed using natural language processing (NLP). We make use of the Pitt Corpus obtained from the Dementia Bank. The dataset contains transcripts of cognitive tests and is used to train and evaluate a variety of machinelearning models, like Bidirectional Encoder Representations from Transformers (BERT), Bio-Clinical BERT, Logistic Regression, Naïve Bayes, and Random Forest. We seek to achieve the capability to accurately detect and classify the problem of Alzheimer’s Disease without the use of physically intrusive methods but instead by incorporating textual features obtained from speech of subjects, that are relatively inexpensive. It has been found that the application of deep learning approaches, such as BERT, bring improvement over traditional methods, indicating the opportunities of using NLP in medical diagnostics. Among the tested models, the BERTbased model registered the best accuracy (85.0%) and AUCROC (0.9126), with high F1-scores for dementia classification (0.90) and moderate performance for controls (0.70).