The study involved chart review, did not include any direct patient involvement, and the data was analyzed anonymously. The study was performed with the approval of the IRBs of the three participating institutions: University of California, Davis Medical Center, Sacramento, California, USA; Children's Hospital Boston, Boston, Massachusetts, USA; Children's Hospital & Research Center Oakland, Oakland, California, USA.

Clinical observations of African American and Hispanic American males and females made at presentation (first visit after determination of hyperglycemia) were obtained from medical records. From 285 charts of African American patients, 58 patients were excluded due to missing data from presentation (n = 50) or unclear diagnosis (n = 8). Data were abstracted from the remaining 227 medical records; 224 of these patients were between the ages of 2 and 18 and were used in the model (Table 1). From 168 charts of Hispanic American patients, 56 patients were excluded due to missing data from presentation (n = 21), unclear diagnosis (n = 28), and diagnosis of diabetes other than T1D or T2D (n = 7). Data were abstracted from the remaining 112 medical records; 105 of these patients, between the ages of 2 and 18, were used for the model (Table 1). Data from children younger than 2 years of age were not used in any analysis that included BMI z-score because BMI measurement in children under 2 may be inaccurate. Presentation dates in the data set spanned two decades, beginning in the 1980s.

The patients were seen at three medical centers: Children's Hospital and Research Center Oakland, Oakland, California, USA; UC Davis Medical Center, Sacramento, California, USA and Children's Hospital Boston, Boston, Massachusetts, USA. The study involved chart review and did not include any direct patient involvement. The study was performed with the approval of the Institutional Review Boards of the three participating institutions. Racial/ethnic status was abstracted from chart data and confirmed by clinician interview.

Data from patients with questionable race/ethnicity, borderline diabetes (e.g., impaired glucose tolerance), diabetes secondary to another condition or treatment (e.g., anti-inflammatory steroid treatment), or known diagnosis of maturity onset diabetes in youth (MODY) were excluded. Because these data were collected retrospectively, with some records many years old, the date of diagnosis and the record of height and weight did not always coincide exactly. Records were used for the study only when height and weight measurements were recorded within one month of presentation. Results of HbA1c analysis were only used if the measurement was made within one month of presentation. Family history of diabetes and presence of acanthosis nigricans (an indicator of high blood insulin levels in type 2 diabetes patients) were considered for use as additional variables in the logistic regression analyses, but data were too sparse to allow their inclusion. Patients younger than two years of age at presentation were excluded from analyses involving BMI z-score.

Established, physician-confirmed diabetes diagnoses were those recorded after at least one year of treatment. The physician used his or her clinical judgment, paired with either direct knowledge of the patient and clinical laboratory data, or thorough examination of both the recorded observational data and the clinical laboratory data to establish the diagnosis. Observational data used in diagnosis included gender, age, height, weight, family history, polyuria, and polydipsia. Clinical laboratory data included some combination of: levels of autoantibodies known to be associated with beta cell destruction (ICA, IAA, GAD65, and IA-2), blood glucose, C-peptide, glycated hemoglobin (HbA1c), blood pH, and diabetes medication. At the time of data abstraction, African American patients in the study had been treated for a mean of 4.9 years (range 1 to 18 years), and Hispanic American patients had been treated for a mean of 3.7 years (range 1 to 10 years). The respective median treatment periods were 4 and 3.5 years. Data from both genders were combined: (1) for determination by race/ethnicity of T2D and T1D rates of decline in age at presentation over the last two decades; (2) for determination by race/ethnicity of mean glycated hemoglobin values in T1D and T2D patients at presentation; and (3) for determination of changes in BMI z-score vs. patient age at presentation.

The logistic regression analysis of three variables – age at presentation, BMI z-score at presentation, and gender – were used to determine sensitivity (correct prediction of T2D), specificity (correct prediction of T1D), and the odds ratios of each variable for risk of T2D.

Graphical representations, statistical significance of differences between data groups, and bivariate linear regression analyses were generated with GraphPad Prism 5 (GraphPad Software, Inc., La Jolla, CA, USA). Pearson's chi-square test with one degree of freedom was used to test for a statistically significant difference in the subgroup distribution of females and males in each type of diabetes. Group means were compared for statistically significant differences by an unpaired t-test with Welch's correction, applied when equal variances cannot be assumed. Statistical differences between slopes of bivariate linear regression lines were determined by an F test.

Multivariate logistic regression analysis was performed with SAS version 9.2 (Cary, North Carolina, USA). From the pool of 224 African American patients, aged 2–18, 150 were randomly selected for the initial analysis. The remaining 74 African American patients served as a replication cohort. From the Hispanic American patient data set, 105 patients, aged 2–18, were submitted for multivariate logistic regression analysis. The statistician was blinded to the patients' diagnoses by physicians until after the computations were complete.

Area under the receiver operating characteristic curves (AUC of ROC curves) was calculated in GraphPad Prism 5 by entering 1) percent sensitivity and 2) 100% minus percent specificity directly into the GraphPad Prism 5 table. The GraphPad program was used only to calculate AUC, not to compute sensitivity and specificity; those values were taken directly from SAS multivariate logistic regression analysis. For each probability computed by SAS, the associated percent specificity was subtracted from 100% to calculate the false positives values. These were entered into GraphPad to create the ROC graph and compute the AUC.

Gender distribution differed significantly (p<0.0001) between the African American T2D patients (72% female) and the African American T1D patients (43% female). However, no significant difference was observed between T2D and T1D gender distribution in the Hispanic American patients (p = 0.42) (Table 1).

BMI is a gender-specific measure of body mass index (BMI) that is normalized for age. The BMI z-score for a given BMI is the number of standard deviations away from the BMI norm for gender and age. For T1D, BMI z-scores were highly variable in both African American and Hispanic American patients, as shown by the large standard deviations of the group means (Table 1). T2D BMI z-scores were not only much higher than T1D scores but also less variable.

Obesity was common among T2D patients at presentation. In addition, a significant percentage of T1D patients were obese (BMI z-score ≥2) at this time. The percentage of obese T1D African American patients was 16% and in Hispanic American patients was 12% (Table 1).

BMI z-scores increased slightly but steadily with an increase in the age at presentation in T1D patients and decreased slightly but steadily in T2D patients (Figure 1). Thus, the T2D scores tended to slightly diminish toward the norm, but were still very high over the onset age range of 2–18 years.

Mean age at onset was lower for females than for males (p<0.04), with the exception of T1D Hispanic American females whose mean age was older than that of T1D males (Table 1).

The decline in age at presentation for T1D and T2D has been shown in other studies [1], [11]. The study reported here demonstrates similar decreases in age at presentation (Figure 2). The decreases in age at presentation illustrate a confounding factor in the differential diagnosis of T2D and T1D and pose a challenge to developing a diagnostic aid for use at presentation. The overlapping of age at presentation between T1D and T2D patients is illustrated graphically in Figure 3. The overlap was particularly pronounced in the age range of 10 to 15 years.

Combining the data from both minority groups, most of the younger T2D patients were females. Of the 51 male T2D patients in the study only one presented younger than 10 years of age. In contrast, 14% of T2D females (12 out of a total of 88 AA and HA females) were under 10 years of age at presentation. This difference was statistically significant (p = 0.031).

The HbA1c data recorded within one month of presentation were available for a subset of 97 African American patients and a subset of 95 Hispanic American patients. No gender differences were observed in HbA1c for either AA or HA patients (data not shown). The mean HbA1c in T2D Hispanic American patients was significantly less than that in T1D (p<0.002) (Table 2). Mean glycated hemoglobin in African American patients was similar in T2D and T1D.

In the abstracted data, several variables differed significantly between T1D and T2D patients, including BMI, gender, C-peptide, anti-GAD autoantibodies, current use of oral diabetes medications, hyperpigmentation, and family history for T2D (data not shown). Of the subset of variables that can be assessed immediately upon patient presentation, only age, gender, and BMI z-score had sufficient data available for use in constructing a predictive model. One hundred fifty of 224 eligible African American patient records were randomly selected for inclusion in the analysis constructing the model. Using a cut point of 0.5, results of regression analysis gave a sensitivity of 92% (correct prediction of T2D) and specificity (correct prediction of T1D) of 89%. Application of the model from this logistic regression analysis to the remaining 74 African American patient records validated the results: sensitivity was 91% and the specificity was 93%.

Application of the model generated from African American data to a cohort of 89 Hispanic American patients produced lower sensitivity and specificity values: 79% and 87%, respectively. The Hispanic American data set was then expanded until the number of eligible patients reached 105. With 105 Hispanic American patients, logistic regression produced a model with sensitivity and specificity of 92% and 90%, respectively. The progress of the model development is summarized in Table 3.

The odds ratios obtained illustrated the importance of female gender as a risk factor for T2D in African Americans and showed that BMI was prominent as a risk factor for both African American and Hispanic Americans. Increasing age at presentation increased the odds of T2D in both African American and Hispanic American patients (Table 4).

The ROC curve is a measure of the utility of the test for prediction of T2D. A ROC area under the curve (AUC) ranges from 0.5 (no predictive value) to 1.0 (perfect predictive value). The AUC of 0.855 (95% CI 0.814–0.896) in the African American patient group indicates that this test for T2D is very useful in differentiating between T2D and T1D. The Hispanic American patient group data yielded a similarly good result: AUC of 0.858 (95% CI 0.808–0.908). The ROC curves are shown in Figure 4.