Stratified analysis of the age-related waist circumference cut-off model for the screening of dysglycemia at zero-cost

Highlights

• Novel artificial intelligence methods can simplify the screening of dysglycemia.

• An age-related waist circumference cut-off model to detect dysglycemia.

• Stratified analysis in sex, race and glycemia group maintaining good performance.

• Good calibration of the model to detect dysglycemia.

• Model is ideal to serve as zero-cost large-scale dysglycemia screening tool.

Abstract

Aims

We perform a stratified analysis of the recently published age-related waist circumference cut-off model to validate its performance in the screening of dysglycemia in the US population.

Methods

We use NHANES data as representative of the US population. Data were subdivided into sex, ethnic and glycemia groups. We evaluate the performance of the model separately in each group through the AUC. We also discuss the calibration of the model.

Results

For the sex-stratified analysis, we obtain AUC = 0.69–0.71 (95% C.I.) for male individuals and AUC = 0.75–0.78 (95% C.I.) for female individuals. The stratified analysis is performed in different ethnic groups, namely “Mexican American”, “Other Hispanic”, “Non-Hispanic White”, “Non-Hispanic Black” and “Other Race – Including Multi-Racial. We obtain, respectively, AUC = 0.74–0.75, AUC = 0.76–0.78, AUC = 0.73–0.75, AUC = 0.74–0.77 and AUC = 0.71–0.73 (95% C.I.). The model achieves AUC = 0.70–0.73 (95% C.I.) in the identification of individuals with prediabetes and AUC = 0.70–0.80 (95% C.I.) in the identification of individuals with diabetes.

Conclusions

The accuracy of the model turns out to be similar in each group considered in the stratified analysis, indicating that the model is suitable to be used as a screening tool for dysglycemia in the US population.

Introduction

The importance of developing screening models to identify subjects with early dysglycemia is an important task to prevent the development of manifest diabetes (Siu and U.S. Preventive Services Task Force, 2015; American Diabetes Association, 2020), to reduce the incidence of several fatal and nonfatal diabetes complication (Lindstrom and Tuomilehto, 2003) and to save the enormous costs of full-blown diabetic disease to the health system (Centers for Disease Control and Prevention, 2020).

Our recently published dysglycemia model (Buccheri et al., 2021) fits into this framework. The model, developed through the support of machine learning (ML) software Brain Project (BP) (Russo, 2016, 2020), predicts the dysglycemia status in a sample of the American adult population through the exclusive use of two zero-cost variables that do not require laboratory tests: waist circumference and age of an individual. As pointed out in Ref. (Buccheri et al., 2021), waist circumference is usually related to the visceral fat of an individual, thus testifying the existence of a strong correlation between obesity and dysglycemia. This result agrees with other important evidence previously published in the literature about genetic predisposition of developing diabetes in patients affected by obesity (Sheikhpour et al., 2020) and the major risk of diabetes and cardiovascular disease in the various obesity types (e.g. central obesity, visceral obesity, etc.) (Miklishanskaya et al., 2021).

Despite its simplicity, the model has similar performance than other complex models previously developed. Therefore, it is ideal to perform large scale screening of dysglycemia in the adult US population (Buccheri et al., 2021). The original model, which in turn consisted in an age-related waist circumference cut-off, was so far validated exclusively on a dataset representative of the general US population. However, its detailed analysis in different sex and ethnic groups, which can be considered intrinsic factors in the genesis of dysglycemia (American Diabetes Association, 2020), is an important step to fully validate the model. For example, it is known that risk to develop dysglycemia in female individuals increases more with age than for male individuals (Huebschmann et al., 2019) as well as certain ethnic groups have a greater genetic predisposition to have an altered blood sugar than others (Gurka et al., 2013).

The paper investigates the stratified performance of the model in different sex and ethnic groups of the US population, which is extremely important in the genesis of a zero-cost and easy-to-use model. We also evaluate the predictive performance individually on prediabetes and diabetes and discuss the calibration of the model. The present work, in turn, complements the findings of our previous work (Buccheri et al., 2021) through a stratified analysis of the model accuracy.