A Systematic Review of Alzheimer's disease detection based on speech and natural language processing

Objectives: The main goal of this review is to intro-duce the work with various conversational datasets containing data from patients suffering from Alzheimer's disease. The basic questions we deal with in the systematic review are: (1) Which datasets are most often used in studies? (2) How does the number of participants in individual datasets vary? (3) What is the representation of individual groups in datasets? Methods: We used databases Scopus, Web of Science and Google Scholar to create the report. Inclusion criteria have been created from key words - voice, speech, Alzheimer detection and natural language processing. We have focused on articles from the beginning of 2019 until now. Articles that did not contain full-text in English were excluded. Results: The review contains a total of 37 studies in which their datasets can be examined. The most commonly used datasets in the articles are ADReSS, PITT and CCC. Many datasets were created for a specific study only, and some have not yet been made public. The range of participants in individual datasets ranges from 30 to 865. The most frequently used ADReSS dataset in the studies contained records of 156 participants. Conclusion: In several cases, the size of the dataset turns out to play an important role, but the overall quality of the dataset was a more significant factor. As a result of a deeper understanding of datasets, we concluded that many factors, such as the age of the participants, gender, or the number of education years, played a significant role as an indicator for Alzheimer's disease prediction.