Semantic analysis in subjects with Alzheimer's disease using Artificial Intelligence

Abstract

In recent years, many investigations have focused on identifying cognitive impairment traits using audio from neuropsychological tests using Artificial Intelligence tools. Unfortunately, the best-performing models require a large amount of data and are not explainable due to their complexity. Given these drawbacks, in this work, we analyze different sentence vectorization models pre-trained on the Pitt Corpus database transcripts to obtain the best model for this task. Likewise, new elements are proposed to analyze semantic features in subjects with Alzheimer's disease. Better results were obtained: 0.8712 accuracy, 0.8729 precision, 0.8712 exhaustiveness, and 0.8709 F1 score using RoBERTa as a model to vectorize sentences and SVM as a classifier in conjunction with the four language features: keywords, circumlocutions, text complexity and similarity to example text. Tests show that the proposed language features increased ranking results, with keywords being the most important.