Characterization of data-driven clusters in diabetes-free adults and their utility for risk stratification of type 2 diabetes

In this study, two independent cohorts were used. Data from the Stockholm Diabetes Prevention Program (SDPP) were used in the main analysis, while data from the Metabolic Syndrome Cohort (MSC) in Mexico were used as a validation dataset (see Additional file 1: Table S1).

Cluster analysis was performed using k-prototypes, an unsupervised partitioning method that divides the dataset based on the variability of categorical and continuous variables [22]. The variables used for the identification of clusters were chosen based on previous studies [14, 16] and included fasting plasma glucose and insulin levels, HOMA2-IR, HOMA2-B, BMI, systolic and diastolic blood pressure, sex, family history of type 2 diabetes, and level of education. Prior to analysis, continuous variables were standardized (mean=0, standard deviation (SD)=1), and extreme outliers, defined as values ≥ ±5 SD, were removed.

Analysis was done independently in the main data (SDPP) and validation data (MSC). We assessed validity based on measures of internal and external validity, internal stability, and visual validation [23]. The gap statistic was used to determine the optimal number of clusters [24] and as a measure of internal validity together with the within clusters sum of squares. External validity was assessed using Cox proportional hazard models to determine whether the clustering process predicts the incidence of type 2 diabetes. The internal stability of the clusters was assessed with the Jaccard similarity index, estimated using 1000 bootstrapped samples. A value greater than 0.75 for each cluster was considered stable [25]. Finally, visual validation was done using box plots and bar charts, as well as uMAP and heatmaps to compare the patterns of the variables used for clustering in each cohort. A detailed description is available as Additional file 3 [22].

The baseline characteristics of the participants, grouped by cluster, are presented as the mean and SD for continuous variables, and as proportions for binary and categorical data (Tables 1 and 2). The risk of type 2 diabetes associated with the resulting clusters was estimated using multivariable Cox proportional hazards models with age as the underlying time variable [26]. The participants in the SDPP study sample were followed from the date of baseline examination to the first recorded date of a new diagnosis of type 2 diabetes, date of death, diagnosis of type 1 diabetes obtained from registers, or until March 31, 2021. The participants in the MSC study sample were followed from the date of baseline examination until the self-reported date of diagnosis of type 2 diabetes, self-reported starting date of glucose-lowering therapy, date of follow-up examination when diagnosed with type 2 diabetes, date of death or until February 28, 2014.

Proportional hazards were assessed visually using log-log plots of survival and predicted Kaplan–Meier survival plots, as well as statistically. Analysis in SDPP was stratified by year of birth using 5-year intervals [26]. Crude and adjusted hazard ratios (HR) are reported. Possible confounders included in the models were sex, self-reported general health status, presence of chronic comorbidities, physical activity level, smoking status, and history of gestational diabetes among women [27].

To evaluate the clinical utility of the clusters, the predictive accuracy and long-term stability of the clusters were assessed and compared to prediabetes in the SDPP data. Predictive accuracy was evaluated using sensitivity, specificity, area under the curve (AUC), and concordance statistics. The long-term stability of the clusters was assessed in the SDPP data from participants who had attended all the study follow-ups using intra-rater agreement measured as Cohen’s kappa and Gwet’s AC1 index [26].

Statistical analyses were performed using Stata version 17 [28]. The clustering algorithm was implemented using a customized k-prototype package in Python 3.7 [29, 30]. The code used can be found as Additional file 3.

Previous studies have used similar methods to investigate the heterogeneity of type 2 diabetes and its complications. Li and collaborators reported three distinct phenotypes of type 2 diabetes using data from digital medical records [13], and Ahlqvist et al. more recently described five clusters of type 2 diabetes using cluster analysis [14]. However, while the clusters reported by Ahlqvist et al. have been replicated in some studies [31, 32], others have failed to reproduce them or found their clinical utility to be limited [33, 34].

Few studies describe clusters before the diagnosis of type 2 diabetes. A study in South Korea by Cho et al. identified six different clusters associated with differences in the prevalence of type 2 diabetes [17], and another by Wagner and collaborators applied cluster analysis to detect phenotypes among individuals at high risk for type 2 diabetes [35].

In all the studies mentioned above, apart from that by Cho et al., analysis required complex data (such as the presence of antibodies to glutamic acid decarboxylase) which are not readily available in most settings. While such data are very valuable, if the aim is to implement cluster analysis widely for the surveillance and prevention of type 2 diabetes, data should be accessible.

On the other hand, important risk factors such as sex and socioeconomic position have not consistently been included, perhaps due to the methodological challenge of using categorical values in cluster analysis. Only the study by Cho et al. included sex, while no previous studies have used indicators of socioeconomic position. In our study, female sex and higher education were overrepresented in the low-risk clusters, indicating the importance of sex and socioeconomic status in the pathophysiological process leading to type 2 diabetes.

The degree of replicability in previous studies is unclear. In the study by Wagner et al., the clusters were replicated using a larger cohort, although they used different variables, assuming that they were conceptually similar, while Ahlqvist et al. and Cho et al. used replication samples derived from the same target population. In contrast, we used comparable data from two independent populations and found comparable patterns.

Additionally, it remains unclear whether different clusters represent different stages in the natural history of type 2 diabetes or etiologically different phenotypes. Although studies have reported the short-term transitions between clusters through time [35, 36], our study is, to the best of our knowledge, the first describing long-term stability.

Our clusters also show important similarities to those reported previously. For example, the phenotypes defined by Ahlqvist et al., severe insulin-deficient diabetes and severe insulin-resistant diabetes, and by Wagner et al. with a prominence of β-cell failure and insulin resistance, resemble our HRBF and HRIR clusters. The studies by Cho et al. and Li et al. describe a subgroup characterized by high blood pressure, like the HRHBP cluster described in this study. The mild obesity-related diabetes cluster described by Ahlqvist et al. and the low-risk obese subgroup described by Wegner et al. are similar to the LRHB cluster, which includes individuals with higher BMI than the population average who had a relatively low risk of type 2 diabetes. However, this was not the most notable characteristic of this group in our study. This difference might be explained by selection bias introduced by clustering individuals with type 2 diabetes or at high risk of type 2 diabetes, among whom high BMI is overrepresented.

The strengths of this study include the use of independent study samples from different countries, which allowed to assess the replicability of our findings. Furthermore, a large and representative sample with a follow-up period of over 20 years (SDPP) was used for the main analysis, which allowed us to estimate the incidence of type 2 diabetes with limited attrition, thanks to the availability of data from regional and national registers.

The replication sample (MSC), in contrast, had a shorter follow-up (3 to 6 years), had a relatively low number of new cases of type 2 diabetes, and was more prone to bias due to sample selection and attrition, which might impact external validation, as the statistical power of the survival analysis to detect an association between cluster assignment and risk of type 2 diabetes might be limited. However, the data demonstrated trends consistent with those of SDPP. Replication using data with longer follow-ups is needed to ensure the external validity of our findings. Likewise, the secondary analysis of the long-term stability on a subset from SDPP might also be biased by loss to follow-up. These results should thus be interpreted with caution.

The selection of risk factors used in the cluster analysis was based in previous literature and included well-known risk factors of type 2 diabetes. Data on biochemical and genetics were not available, and it is unclear whether it could have affected the results of the cluster analysis. Nevertheless, such information is not usually available in most settings.

Incorrect classification of type 2 diabetes as type 1 or other sorts of diabetes is a common problem in epidemiological studies [37]. In SDPP, data from clinical registries allowed us to identify individuals with other types of diabetes directly. In the MSC study, we limited the sample to individuals with a diagnosis after 30 years of age to minimize the risk of including individuals with type 1 diabetes. However, neither method completely eliminates the risk of misclassification bias.

From a public health perspective, data-driven risk stratification could lead to better-targeted interventions and resource utilization for the prevention of type 2 diabetes. The high-risk clusters described in this study define a group with high sensitivity and specificity that contains about one-third of the study population and captures a large majority of the cases of type 2 diabetes, which might have important implications for public health and clinical practice. However, questions remain regarding the practicality of using cluster analysis, its utility for guiding public health policy and clinical decisions, and the underlying physiological mechanisms driving the different phenotypes.

The first step towards the practical implementation of data-driven stratification is assigning individuals to the most appropriate cluster. This can be accomplished using data available in health registers or electronic medical records. The observed differences between the Swedish and Mexican cohorts indicate that clustering should be population specific, although our results show that this is likely to result in very similar phenotypes.

Clusters could also be used to investigate the effect of different interventions to prevent type 2 diabetes. For example, the relatively high blood pressure in the HRHBP cluster suggests that blood pressure control could be an important part of the intervention for this group of individuals. The high fasting glucose and insulin levels of individuals in the HRBF cluster resemble the β-cell dysfunction characteristic of type 2 diabetes; thus, pharmacological interventions to regulate β-cell function might be effective. On the other hand, intensive lifestyle interventions might have the greatest effects in individuals in the HRIR cluster, who were characterized by the highest BMI and tendencies towards insulin resistance.

Likewise, studies looking into the association between different clusters and complications of type 2 diabetes may aid clinical decision-making in patients with type 2 diabetes and lead to new insights into the natural history and pathophysiology of the disease.

Finally, to better understand the physiological mechanisms underlying the differences between clusters, studies of the environmental, social, genetic, and biochemical determinants of the different clusters are needed.