OBJECTIVES: Conventionally, diabetes is classified into Type-1 and-2 diabetes, but this classification offers limited insights into underlying mechanisms and heterogeneity within these types. Recently, a variety of novel diabetic patient clusters have been identified, using unsupervised machine-learning approaches. We aimed to identify all novel data-driven clusters of diabetic patients and to summarize patient characteristics and incidence of diabetic complications in each cluster.

METHODS: A systematic review (PROSPERO CRD42023406046) of English-language journal articles was undertaken in PubMed, Embase, Scopus, and Web of Science from 1st January 2017 to 15th February 2023. References cited by included studies will be taken into account to identify publications before 2017 and potential missing literature. All studies were screened independently by two authors. Data extraction was conducted by one author and checked by another, using a pre-defined form.

RESULTS: One hundred and sixteen studies were identified, among which fifty-two were related to of Ahlqvist (2018), including replication (36/52), validation (10/52) and both (6/52). The other 65 studies proposed new cluster definitions. Meta-analyses of patient characteristics (including age at diagnosis, BMI, HbA1c, HOMA2-B, HOMA2-IR) and incidence or relative risk of diabetic complications were performed for these 51 studies. The most common complications were nephropathy (15/52), retinopathy (15/52), neuropathy (12/52) and cardiovascular diseases (12/52).

CONCLUSIONS: This review provided an overview of recent studies proposing novel diabetes subgroups based on clustering analysis. The cluster strategy proposed by Ahlqvist (2018) is the cornerstone of various follow-up studies. We found differences in data sources, patient population, algorithms and clustering variables, and outcomes of interests. The meta-analysis results provided summary estimates of centroids in each cluster, which enabled more reliable cluster classification in implementation.