WHAT IS ALREADY KNOWN ON THIS TOPIC

Previous research indicates that diabetes risk prediction models incorporating body mass index (BMI), fasting plasma glucose (FPG), and HbA1c demonstrate effective identification of high-risk individuals, while the impact of adding HOMA-IR and HOMA-β and substituting visceral fat area (VFA) for BMI remains unknown.

WHAT THIS STUDY ADDS

This study expands current knowledge by examining the influence of adding HOMA-IR, HOMA-β, and substituting VFA for BMI in diabetes prediction models.

The findings indicate that including these factors in models with BMI, FPG, and HbA1c does not substantially enhance predictive accuracy.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICYIntroduction

Type 2 diabetes mellitus is a major public health problem, which affected around 6% of the world’s population in 2017 and was projected to increase to over 7% in 2030.1 Japan is one of the top 10 countries which have the highest number of adults with diabetes.2 Early identification of people at increased risk of developing diabetes is critical for taking measures to delay or prevent the condition.

Numerous risk prediction models for diabetes based on different sets of predictors have been developed worldwide.3 We previously developed a highly accurate prediction model including both HbA1c and fasting plasma glucose (FPG).4 A recent study of US community-dwelling adults reported that the predictive performance for diabetes significantly improved with the addition of HOMA-IR, an insulin resistance (IR) marker, to the model including HbA1c.5 Another study also showed that adding IR markers (fasting insulin and C-peptide) improved diabetes risk prediction beyond FPG and other risk factors (HbA1c was not included).6 However, it remains uncertain whether the performance of risk models which already included both HbA1c and FPG could be further improved by adding IR. Similarly, it is also unclear whether adding HOMA-β could improve the model performance. This information is more relevant when predicting the risk of diabetes in populations where insulin deficiency plays a more important role in the development of diabetes than IR, such as the Japanese population.7 8 In addition, visceral fat area (VFA) is reported to be a better predictor for diabetes than body mass index (BMI).9–11 It is thus anticipated that risk prediction models perform better if VFA is substituted for BMI. Here, we tested whether the performance of diabetes risk prediction models can be improved by adding HOMA-IR and HOMA-β and replacing BMI with VFA among the Japanese population.

Methods

The present study was conducted among workers of a large company participating in the Japan Epidemiology Collaboration on Occupational Health (J-ECOH) study, details of which have been described elsewhere.4 We analyzed data of annual comprehensive health check-up (Ningen Dock) between fiscal year 2006 and 2020, including computed tomography (CT)-based VFA, fasting insulin, as well as information on lifestyles and medical history. Prior to data collection, the J-ECOH study was announced in each participating company using posters. Participants were allowed to refuse participation (opt-out) and did not provide verbal or written informed consent.

Analytic cohort

We regarded the fiscal year 2006 as the baseline of the analytic cohort. If the 2006 data were not available, the 2007 data were used. At baseline, a total of 6597 participants (aged 30 and over) underwent the comprehensive health checkup and had data for predictors of diabetes including CT-based VFA, HOMA-IR, HOMA-β, BMI, smoking, hypertension, and dyslipidemia. Of these, we excluded individuals who met any of the following criteria: ongoing anti-diabetic drug therapy, self-reported history of diabetes, FPG levels of ≥126 mg/dL, HbA1c levels of ≥6.5% (48 mmol/mol), or a history of cancer or cardiovascular disease (n=700). Of the remaining 5897 participants, we excluded those who did not attend any subsequent health checkup or who attended but received neither glucose measurement nor HbA1c measurement (n=319), leaving 5578 for the present analysis. Participants included in the final sample were on average younger and had a higher prevalence of smoking, while exhibiting a lower prevalence of hypertension compared with those who did not attend subsequent health checkups or had incomplete glucose/HbA1c measurements. No significant differences were observed in terms of blood glucose, VFA, HOMA-IR, and HOMA-β (online supplemental table 1).

Outcome

Incident diabetes was ascertained using annual health examination data collected after the baseline until fiscal year 2020. Type 2 diabetes mellitus was defined as meeting at least one of the following criteria, in accordance with the criteria of the American Diabetes Association12: a FPG level of ≥126 mg/dL, a random plasma glucose level of ≥200 mg/dL, an HbA1c level of ≥6.5% (48 mmol/mol), or self-report of currently being under medical treatment (medication or lifestyle modification) for diabetes.

Assessment of predictor variables

Referring to our previous prediction models for diabetes,4 the following predictor variables were included: sex, age, BMI (kg/m2), current smoking (yes or no), hypertension (yes or no), dyslipidemia (yes or no), FPG (<100, 100–<110 or 110–<126 mg/dL), HbA1c (≤5.5% (37 mmol/mol), 5.6–5.9% (38–41 mmol/mol), or 6.0–6.4% (42–47 mmol/mol)), VFA, HOMA-IR, and HOMA-β. Body weight and height were measured using a scale while the participant wore light clothes and no shoes. BMI was calculated as the weight in kilograms divided by the squared height in meters. Smoking status was identified by a self-administered questionnaire. Blood pressure was measured with the participants in a sitting position using an automatic blood pressure monitor. Hypertension was defined as systolic blood pressure ≥140 mm Hg, diastolic blood pressure ≥90 mm Hg, or receiving medical treatment for hypertension.13 Dyslipidemia was defined as triglyceride level ≥150 mg/dL, low-density lipoprotein-cholesterol level ≥140 mg/dL, high-density lipoprotein-cholesterol level <40 mg/dL, or receiving medical treatment for dyslipidemia.14 VFA was measured at the umbilical level using CT scans as previously described.15 Fasting glucose levels were assayed using the glucose electrode technique. HbA1c was measured using high-performance liquid chromatography. The HbA1c values collected before the 2013 fiscal year were assessed using the Japan Diabetes Society criteria and were converted to the National Glycohemoglobin Standardization Program value using the following formula: HbA1c (%) = 1.02 × HbA1c (Japan Diabetes Society) (%) + 0.25%. Serum immunoreactive insulin (μU/mL) was measured using an immunoenzymatic method. HOMA-IR and HOMA-β were calculated using the formulas: fasting glucose (mmol/L)*fasting insulin (μU/mL)/22.5 and 360 * fasting insulin (μU/mL) / (fasting glucose (mg/dL) − 63), respectively.

Statistical analysis

The characteristics of the study participants at baseline were described as means for continuous variables and percentages for categorical variables. χ2 tests for categorical variables or t-tests for continuous variables were used to examine differences in baseline characteristics between individuals with incident diabetes and those who did not develop it.

Person-time was calculated from the date of the baseline examination to the date of first heath examination when diabetes was identified or to the date of the last health checkup, whichever occurred first. The Cox proportional hazards regression analysis was used to develop the risk models for diabetes, as in our previous study.4 We initially conducted a baseline model (model 1) including age, sex, BMI, smoking, dyslipidemia, hypertension, and HbA1c. Subsequently, we developed another four models: model 2, predictors in model 1 plus FPG; model 3, predictors in model 1 plus HOMA-IR and HOMA-β; model 4, predictors in model 1 plus FPG, HOMA-IR, and HOMA-β; model 5, replaced BMI with VFA in model 2. All models demonstrated a low level of multicollinearity, specifically with a maximum variance inflation factor of around 3.5 for both HOMA-IR and HOMA-β in model 4.

We compared the performance of models 2–5 with model 1 by examining measures of discrimination and calibration. Discrimination is quantified by calculating the area under receiver operating characteristic curve (AUROC) for 10-year diabetes risk. In addition, integrated discrimination improvement (IDI) and continuous net reclassification improvement (NRI) were computed to show the improved performance (if any) of models 2–5 compared with model 1. In our study, the events per variable number is approximately 90 (943/11), significantly exceeding the recommended minimum of 10 to prevent overfitting. Therefore, the likelihood of overfitting is low. We further assessed calibration in two ways: visually by plotting the predicted 10-year risk of diabetes versus the observed risk in a calibration plot and performing the Greenwood-Nam-D' Agostino goodness-of fit test.

In sensitivity analyses, we reanalyzed the data after excluding women and participants with less than 1 year of follow-up, analyzed data by obesity status in men, and substituted HOMA-IR and HOMA-β with fasting insulin. All statistical analyses were performed using SAS V.9.4 (SAS Institute). A two-sided p<0.05 was considered statistically significant.

Results

Table 1 shows the baseline characteristics of the participants. Overall, the mean age of participants was 48.4 (8.2) years, and majority of participants were men (5281 men and 297 women). Over a median follow-up of 10 years (range, 0.5–15 years), 943 participants developed diabetes. The incidence rate of diabetes was 18.4 per 1000 person-years. Individuals who developed diabetes had higher levels of BMI, FPG, HbA1c, VFA, HOMA-IR, and HOMA-β at baseline.

Table 1

Baseline characteristics of study participants, overall and by incident diabetes status

Table 2 shows the coefficients associated with each predictor of diabetes. As shown in models 3 and 4, HOMA-IR and HOMA-β were significantly associated with the development of diabetes. In model 5, individuals in the highest tertile of VFA had a significantly higher risk of developing diabetes compared with that in the lowest tertile.

Table 2

Multivariate regression coefficients (SEs) of diabetes risk prediction models

As shown in figure 1, the baseline model (model 1) had an AUROC of 0.79 (95% CI 0.78, 0.81). The addition of FPG to model 1 increased the value of AUROC (model 2, 0.84 (0.83, 0.85)). Compared with model 1, model 2 also significantly improved the risk reclassification and discrimination, with an NRI index of 0.61 (0.56, 0.70) and IDI index of 0.09 (0.08, 0.10) (table 3). As for model 3, the addition of HOMA-IR and HOMA-β to model 1 resulted in a relatively small improvement. The performance of model 4 is similar to that of model 2. As shown in figure 1 and table 3, the values of AUROC, NRI, and IDI for model 4 were 0.84 (0.83, 0.85), 0.67 (0.60, 0.75), and 0.09 (0.08, 0.11), respectively. There was no material improvement in prediction performance after replacing BMI in model 2 with VFA (model 5).

Figure 1

Receiver operating characteristic curves for each risk model in predicting 10-year diabetes risk. Model 1: age, sex, BMI, smoking, dyslipidemia, hypertension, and HbA1c; model 2, predictors in model 1 plus FPG; model 3, predictors in model 1 plus HOMA-IR and HOMA-β; model 4, predictors in model 1 plus FPG, HOMA-IR, and HOMA-β; model 5, replaced BMI with visceral fat area in model 2. AUROC, area under receiver operating characteristic curve; BMI, body mass index; FPG, fasting plasma glucose.

Table 3

Discriminative ability of prediction models in comparison with the baseline model

Figure 2 shows the agreement between observed outcomes and predictions in deciles of the predicted risk. No obvious difference between observations and predictions was observed in all models. Compared with model 1, a larger spread between deciles of predicted risk was observed in models 2, 3, 4, and 5, indicating a better discriminating model.

Figure 2

Calibration plot for type 2 diabetes risk models, by deciles of predicted risk. Model 1: age, sex, BMI, smoking, dyslipidemia, hypertension, and HbA1c; model 2, predictors in model 1 plus FPG; model 3, predictors in model 1 plus HOMA-IR and HOMA-β; model 4, predictors in model 1 plus FPG, HOMA-IR, and HOMA-β; model 5, replaced BMI with visceral fat area in model 2. BMI, body mass index; FPG, fasting plasma glucose.

The study findings were supported by sensitivity analyses. Model 2, which included both HbA1c and FPG, consistently demonstrated strong predictive performance for diabetes risk. Nevertheless, making additional modifications, such as adding HOMA-IR and HOMA-β (or fasting insulin) or substituting BMI with VFA, did not result in significant improvements in predictive accuracy (online supplemental tables 2-4).

Discussion

In this cohort study of a working population, we found limited evidence to suggest that HOMA-IR, HOMA-β, and VFA significantly improve the performance of diabetes risk prediction models. To our knowledge, this is the first study to comprehensively assess the value of HOMA-IR, HOMA-β, and VFA in diabetes risk prediction.

Our findings align with previous studies, indicating that the risk prediction models including both HbA1c and FPG can predict the risk for developing diabetes with satisfactory accuracy.4 16 17 Our results extend these findings by demonstrating that further adding HOMA-IR and HOMA-β to the model does not materially improve the prediction performance. IR and defects in pancreatic beta cells are the two major pathophysiological abnormalities that underlie diabetes, and beta-cell dysfunction makes the main contribution to the development of diabetes in Japanese.18 After adding HOMA-IR and HOMA-β to model 1, model 3 showed a better performance than model 1, but the extent of improvement was much less compared with model 2 which included both HbA1c and FPG. Our findings suggest that HOMA-IR and HOMA-β are not potent predictors for diabetes, which can be explained, at least partly, by the beta-cell compensation for IR that occurred before the onset of diabetes. As shown in our study, people who developed diabetes had higher baseline levels of HOMA-β than people without diabetes. Additionally, the superior performance of models incorporating FPG over HOMA-IR and HOMA-β in predicting diabetes risk may be linked to the essential role of FPG in diabetes diagnosis. With a large sample size, long follow-up period, and availability of data on both HbA1c and FPG, our study provided evidence that a prediction model including both HbA1c and FPG is well-suited for identifying individuals at an increased risk of diabetes, while the added benefit of incorporating HOMA-IR and HOMA-β remains limited.

Unlike general obesity, visceral obesity is thought to be directly related to the pathogenesis of diabetes, as visceral fat secretes proteins (eg, visfatin and interleukin 6) and hormones (eg, leptin and resistin) that can cause inflammation and IR, which in turn increase the risk of developing diabetes.19 20 The independent association between visceral fat and diabetes has been reported in cross-sectional and cohort studies,9–11 21 22 some of which reported a stronger association with visceral fat than with BMI.9–11 In the present study, however, we observed no obvious improvement in model performance with substitution of VFA for BMI, suggesting that BMI, a simple index of obesity, can perform as well as VFA for predicting diabetes. After removing the strong predictors HbA1c and FPG (both were also used for diabetes diagnosis), BMI still showed similar predictive capacities for diabetes compared with VFA (AUROC: 0.68 vs 0.67, respectively). The NRI and IDI were 0.12 and 0.005, respectively, when compared with the model that included VFA. To our knowledge, this is the first study to compare the performance of BMI and VFA in diabetes prediction models. Future studies are needed to confirm our findings.

The strengths of this study include the long-term observation, sufficient cases of diabetes, use of blood glucose and HbA1c for diagnosing diabetes, and the availability of HOMA-IR and VFA. Our study also has some limitations. First, the majority of subjects in this study were men, limiting the generalizability of the findings primarily to male subjects, particularly Japanese male workers, and also due to a relatively small number of female participants, we were unable to examine whether model performance differs by sex. Second, HOMA is used for assessing beta-cell function and IR in our study. It is unclear whether the results would be different if it was assessed by other methods such as the hyperinsulinemic-euglycemic clamp, Matsuda index, and the insulin secretion rate. Third, there is potential for intervention bias owing to the provision of Specific Health Checkups and Specific Health Guidance to insured individuals aged over 40 with visceral fat accumulation in Japan, which could have influenced the observed development of diabetes in our study population. Last, the loss of follow-up data from individuals who retired due to age has implications for the generalizability of our risk prediction model. Therefore, caution should be exercised when applying our risk prediction model to older age groups and other populations.

Conclusions

Our study shows that a model with BMI, FPG, and HbA1c effectively identifies those at high diabetes risk. However, adding HOMA-IR, HOMA-β, or replacing BMI with VFA does not significantly improve the model. In clinical and public health practice, consider using readily available BMI, FPG, and HbA1c for diabetes risk prediction, while the incorporation of HOMA-IR, HOMA-β, and VFA, which are costlier and more challenging to measure, may not be as advantageous. Further research is required for confirmation.

Data availability statement

Data are available upon reasonable request. The datasets are not publicly available due to privacy/ethical reasons but are available from Dr Tetsuya Mizoue on reasonable request.

Ethics statementsPatient consent for publication

Not applicable.

Ethics approval

This study involves human participants and was approved by the Ethics Committee of the National Center for Global Health and Medicine, Japan (NCGM-G-001140). Prior to data collection, the J-ECOH study was announced in each participating company using posters. Participants were allowed to refuse participation (opt-out) and did not provide verbal or written informed consent.

Acknowledgments

We acknowledge scientific advice for implementing J-ECOH Study from Dr Toshiteru Okubo (National Institute of Occupational Safety and Health, Japan), implementation of data management from Ms Maki Konishi (National Center for Global Health and Medicine), and support for administrative issues from Ms Rika Osawa (National Center for Global Health and Medicine).