Clinical applications of artificial intelligence in diabetes management: A bibliometric analysis and comprehensive review

Background: Diabetes is one of the most common pathologies today and has become a constant problem in public health worldwide. Objective: The purpose of this study is to deepen the knowledge of the most applied type of diabetes, the methods to test the performance of the model, the most used techniques with better accuracy, the metrics and the programming languages used in the detection of diabetes. To give a more holistic view of this area of health, this article presents a systematic review of literature (SLR) to analyze artificial intelligence techniques in the detection of diabetes. Methods: A comprehensive search of five databases: MDPI, ScienceDirect, Scopus, Taylor and Francis, and Web of Science using keywords was carried out, which identified 70 articles between 2017 and 2023. This article followed the PRISMA methodology, considering inclusion and exclusion criteria, with the aim of synthesizing the findings related to various aspects: types of diabetes, methods to test model performance, techniques, metrics and programming languages; In addition, new challenges and future work were proposed. Results: The type of diabetes most applied was Type-2 Diabetes Mellitus, the methods to test model performance most used were cross validation with 10 and 5 k-folds; on the other hand, the most efficient technique was Random Forest, the same that obtained the best accuracy, the essential metrics to determine the effectiveness of the model were Accuracy and F1-Score and finally the most used programming language to develop the models was Python. Conclusions: This research provides scientific evidence on how machine learning and deep learning techniques can improve diabetes detection. Therefore, this article seeks to contribute to the field of health, specifically to the detection, management and treatment of diabetes using artificial intelligence algorithms as a competitive advantage. It also provides information to make more accurate clinical decisions, for which we identified gaps in the literature related to diabetes screening.