Patient characteristics (Fig. 1)
Parental history of diabetes was more common in MODY patients (90% had at least one parent affected vs 61% in type 2 diabetes and 19% in type 1 diabetes) (Fig. 1). The 10% of MODY cases without a parent affected is most likely to reflect cases where the parent has undiagnosed diabetes at referral or, in a minority of cases, where the mutation is de novo. Patients with GCK MODY were more likely to be treated with diet alone (78%), whereas HNF1A/4A MODY and type 2 diabetic patients were most likely to be treated with OHAs and/or insulin (83% and 91%, respectively). By definition, all patients with type 1 diabetes were treated with insulin. There was a higher proportion of female patients in the MODY groups (65% in GCK MODY, 72% in HNF1A/4A MODY, 53% in type 1 diabetes, 50% in type 2 diabetes).
Figure 1 shows density plots of the continuous characteristics of the four subtypes of diabetes. Type 2 diabetic patients were diagnosed at an older age compared with the other three types of diabetes (mean 30.7 vs 16.8 years, p < 0.001). Patients with type 1 diabetes had the highest HbA1c (mean 9.1% (76 mmol/mol) vs 7.4% (57 mmol/mol) in other types of diabetes, p < 0.001), reflecting poorer glycaemic control. Type 2 diabetic patients tended to be more obese compared with the other three subtypes of diabetes (mean BMI [children adjusted to adult values] 33.1 vs 25.1 kg/m2, respectively, p < 0.0001).
Type 2 diabetes vs MODY models for patients who are not treated with insulin within 6 months of diagnosis
The best model for discriminating type 2 diabetes from MODY was found using logistic regression. Further details of the analysis are provided in ESM Results 1. Table 1 shows the characteristics and associated beta-coefficients and ORs from this model. The strongest predictor of MODY was age at diagnosis, with younger patients being more likely to have MODY (OR for MODY 0.73, [95% CI 0.68, 0.78] for every year increase in age at diagnosis). Other variables associated with higher odds of MODY were being slimmer (as reflected by BMI), having a parent affected with diabetes, not being treated with insulin or OHAs, having better glycaemic control (as reflected by HbA1c), female sex and current age (see Table 1: T2D vs MODY model). It made little difference whether unadjusted BMI, or BMI with child values converted to the adult equivalent, was entered into the model. For ease of use in practice, unadjusted BMI was chosen in the final model.
Figure 2a shows a boxplot of the fitted probabilities for MODY from this logistic regression model for both the MODY and type 2 diabetic patients. The majority of patients with type 2 diabetes have low predicted probabilities for MODY, and the majority of MODY patients have high predicted probabilities. The c-statistic (Fig. 2b) was 0.98, indicating that the model showed excellent discrimination between MODY and type 2 diabetes. Table 2 shows the sensitivities, specificities and corresponding likelihood ratios for MODY using different probability cut-offs in the model. The probability cut-off maximising both sensitivity (92%) and specificity (95%) was found to be 60% (Table 2: T2D vs MODY model).
The model worked well when comparing type 2 diabetes with HNF1A/4A and GCK MODY separately (ROC AUC = 0.974 and 0.990, respectively). We also examined the subgroup of patients diagnosed between 25 and 35 years with HNF1A/4A MODY, as this is the most difficult group to identify, and found that the model still worked well in separating HNF1A/4A MODY patients from those with type 2 diabetes (ROC AUC = 0.915).
The prediction error from jack-knife cross-validation was 5%, suggesting that approximately 95% of cases would be correctly classified given this model. In the external test dataset, the model performed well, with a c-statistic of 0.94. A 60% probability cut-off provided 93% specificity and 87% sensitivity for MODY.
Type 1 diabetes vs MODY models for patients who are treated with insulin within 6 months of diagnosis
The best model for discriminating type 1 diabetes from MODY was also found using logistic regression. Further details of the analysis are provided in ESM Results 1. The strongest predictor of MODY in this model was having a parent with diabetes, with those with at least one parent affected having odds of MODY approximately 23 times higher (95% CI 12.1, 46.9) than those without a parent affected. Other significant predictors of MODY were lower HbA1c, female sex, older age at diagnosis and current age (see Table 1: T1D vs MODY model).
The model showed good performance with type 1 diabetic patients having low predicted probabilities, and the majority of MODY patients having high predicted probabilities for MODY (Fig. 2c). The c-statistic was 0.95 (Fig. 2d). The sensitivities, specificities and corresponding likelihood ratios for MODY using different probability cut-offs in the model are shown in Table 2. The probability cut-off maximising both sensitivity (87%) and specificity (88%) in this model was found to be 40% (Table 2: T1D vs MODY model).
The model worked well when considering type 1 diabetes with HNF1A/4A and GCK MODY separately (ROC AUC = 0.934 and 0.976, respectively), and when comparing HNF1A/4A MODY with type 1 diabetes in only those diagnosed between 25 and 35 years (ROC AUC = 0.952).
The prediction error from jack-knife cross-validation was 9.2%, suggesting that approximately 90% of cases would be correctly classified. In the external test dataset, the model performed well, with an ROC AUC of 0.95. A 40% probability cut-off provided 93% specificity and 82% sensitivity for MODY.
Comparison with traditional MODY criteria
Standard criteria for MODY (age at diagnosis <25 years and parent affected with diabetes) were specific, correctly excluding MODY in 543/597 (91%) type 1 diabetes or type 2 diabetes cases, but had lower sensitivity, picking up only 425/594 (72%) MODY cases. Using the optimal cut-offs identified, the prediction models were more sensitive at picking up proven MODY cases (539/594 [91% sensitivity]), with similar specificity (560/597 [94%]), correctly classifying more patients overall (92% vs 81%, p < 0.0001).
The best linear discriminant and classification tree models are described in the ESM (ESM Results 2, ESM Table 1 and ESM Figure 1).
Positive predictive values (post-test probabilities) for MODY
The positive predictive values and negative predictive values for MODY, adjusting for prior probabilities of 0.7% and 4.6% (for the type 1 diabetes and type 2 diabetes comparisons, respectively) are presented in Table 2.
Compared with the type 2 diabetes vs MODY model the positive predictive values obtained in the type 1 diabetes vs MODY model were more modest, reflecting the lower prior probabilities.
Online clinical prediction calculator
The equations obtained from both logistic regression models were used to produce a web-based version of the models now available at www.diabetesgenes.org. The clinical features of a particular patient diagnosed ≤35 years can be entered into this online form. For patients who are treated with insulin within 6 months of diagnosis, the type 1 diabetes vs MODY equation will be applied to their clinical features, otherwise the type 2 diabetes vs MODY equation will be applied, and the positive predictive value for MODY calculated.