Clinical Prediction Models Combining Routine Clinical Measures Have High Accuracy in Identifying Youth-Onset Type 2 Diabetes Defined by Maintained Endogenous Insulin Secretion: The SEARCH for Diabetes in Youth Study Free

Type 2 diabetes (T2D) is increasingly common in children and young adults, with the prevalence in U.S. youth aged 10–19 years doubling between 2001 and 2017 (1). Identification of youth-onset T2D can be particularly challenging, as features of type 1 diabetes (T1D) and T2D overlap. For example, many children with diabetes and obesity have T1D (2). Islet autoantibody testing may assist classification, especially at diagnosis, but these tests have imperfect specificity and sensitivity, meaning that a positive autoantibody may not confirm T1D, and negative autoantibodies may not exclude this condition (3,–,5). While C-peptide may assist classification, it is often retained at diagnosis in individuals with T1D and obesity, limiting utility at diagnosis (6,7).

The glycemic treatment requirements of T1D differ markedly from T2D. These differences are mainly driven by the development of severe insulin deficiency in T1D but not in T2D (6). In T1D, the severe insulin deficiency that occurs in most people living with this condition leads to absolute insulin requirement, need for physiological insulin replacement, high glucose variability and hypoglycemia risk, and lack of glycemic benefit from most noninsulin glucose-lowering therapies (6,8–11). Within individuals with long-standing diabetes, assessment of insulin secretion, using C-peptide, has been shown to predict successful insulin withdrawal, glucose variability and hypoglycemia risk, and response to prandial insulin and noninsulin glucose-lowering therapies, regardless of clinician-diagnosed diabetes subtype (6,12–18). C-peptide levels greater than ∼250 pmol/L (0.75 ng/mL) fasting or 600 pmol/L (1.8 ng/mL) stimulated have been suggested to identify individuals with T2D and its treatment requirements, and this threshold has been incorporated into international guidelines (10,19).

No single feature or biomarker at diagnosis can robustly identify diabetes subtype and future retained endogenous insulin. The interpretation of individual features or biomarker tests, such as islet autoantibodies, depends on the prior likelihood, informed by other features. For example, islet autoantibodies will have a very low negative predictive value in a child with classical features of T1D but a high negative predictive value where the clinical features also support an alternative diagnosis. Conversely, a single positive islet autoantibody may have moderate positive predictive value where T1D is unlikely (5,20). One way to address this issue is to combine different clinical features and biomarkers using prediction model approaches. In adult-onset diabetes, prediction models with high accuracy have been developed to differentiate T1D and T2D defined by early insulin requirement and long-standing persistence of C-peptide (21,22). The same models have also shown high accuracy for classification of diabetes defined genetically and by pancreatic histology (23,24). However, no clinical models are available for retained C-peptide or classification in youth-onset diabetes. We therefore developed clinical prediction models to combine clinical features and biomarkers, assessed close to diabetes diagnosis, to identify individuals with youth-onset diabetes who will retain high levels of endogenous insulin secretion consistent with T2D.

We developed prediction models for retained endogenous insulin, consistent with T2D and noninsulin requirement (fasting C-peptide >0.75 ng/mL or 250 pmol/L 3–10 years after diabetes diagnosis) in participants in the SEARCH for Diabetes in Youth (SEARCH) study, a multicenter prospective study (25,26).

SEARCH performed population-based ascertainment of youth with incident diabetes diagnosed before age 20 from 2002 through 2020 at five centers in the U.S. as previously described (25,26). Participants with recently diagnosed diabetes in 2002–2006 and 2008 participated in a baseline research visit. Those who had a baseline research visit were invited to participate in a follow-up visit (at a median diabetes duration of 8 years), with serial follow-up visits undertaken in a subset of participants. At each research visit, fasting blood samples (≥8-h fast) were obtained. Diabetes type determined by the participant’s health care provider was recorded and is referred to as provider type. The SEARCH study was reviewed and approved by the local institutional review boards that had jurisdiction over the local study population, and all participants provided informed consent and/or assent. For this analysis, participants were included if C-peptide measurement was available between 3 and 10 years (inclusive) after diabetes diagnosis and the following information from the baseline visit was available in the study data set: age at diagnosis, BMI, waist circumference, sex, HDL cholesterol (HDL-C), GAD, and IA2 islet autoantibody status. Those with monogenic diabetes confirmed by genetic testing were excluded from this analysis.

Substantial retained insulin secretion was defined as a fasting C-peptide >250 pmol/L (>0.75 ng/mL) between ≥3 and ≤10 years’ diabetes duration. When C-peptide was <250 pmol/L (<0.75 ng/mL), concurrent glucose of ≥72 mg/dL (4 mmol/L) was required for inclusion of the result. Where fasting C-peptide was not available, a random nonfasting C-peptide value was used (<0.5% of included measures), converted to the equivalent fasting value by dividing by 2.5 (6).

Model predictor variables were prespecified based on previously reported associations with differentiating T1D and T2D, availability in the study, and lack of need for fasting blood collection. Models included sex, BMI, waist circumference, age of diabetes diagnosis, HDL-C, race and ethnicity, islet autoantibodies to GAD65 (GADA) and IA-2 antigen (IA-2A), and a 67-single nucleotide polymorphism (SNP) genetic risk score for genetic susceptibility to T1D (T1DGRS2) (27). A 397-SNP genetic risk score for T2D (28), and zinc transporter (ZnT8) islet autoantibodies (both available in only a subset of participants) were assessed for improvement in prediction in addition to the above features. Model predictors were assessed at the baseline visit. Where individual parameters were missing at baseline (<0.01% of included parameters), the closest available measure to diagnosis from later study visits was used. Waist circumference was assessed using the National Health and Nutrition Examination Survey method (uppermost lateral border of the ilium). Race and ethnicity were self-reported using 2000 U.S. census questions and defined as four groups based on prevalence in the SEARCH study: Black, White-Hispanic, non-Hispanic White, and other (predominantly Asian and Native American). To provide the greatest flexibility for clinical and research use, models were developed and tested with and without variables that are not routinely measured (islet autoantibodies and genetic risk scores).

Diabetes autoantibodies (GADA, IA-2A and ZnT8), HDL-C, HbA_1c, fasting C-peptide, and glucose were assessed for fasting blood samples at the Northwest Lipid Metabolism and Diabetes Research Laboratories, University of Washington, the central laboratory for SEARCH. The cutoff values for islet autoantibody positivity were 33 National Institute of Diabetes and Digestive and Kidney Disease (NIDDK) units (NIDDKU)/mL for GADA, 5 NIDDKU/mL for IA-2A, and 0.02 NIDDKU/mL for ZnT8 autoantibody. Serum C-peptide concentration was determined by a two-site immunoenzymetric assay (Tosoh 1800; Tosoh Bioscience).

Genotyping and assessment of genetic risk scores in SEARCH have been recently described (28). We calculated two previously published T1DGRS, including a 30-SNP score used in many studies (T1DGRS-1 [29]) and a more recent 67-SNP score with a greater diversity of HLA variation, including 18 interactions between major HLA class II alleles and additional independent variants (T1DGRS-2 [27]). We assessed the T1DGRS-2 in the primary analysis. We also assessed a T2D genetic risk score (T2DGRS) generated from summary statistics of a previously published comprehensive genome-wide association study in individuals of European ancestry (28,30).

We determined sample size using the approach described Riley et al. (31) and the pmsampsize package in R (32). For an outcome prevalence of 15%, eight model parameters, and assuming a model optimism-adjusted C-statistic of >0.8, the minimum sample size required for small overfitting (<10% expected shrinkage of predictor effects), small mean average prediction error, and small margin of error around estimation of the outcome probability (±0.05) was 438. The SEARCH cohort contains 1,918 participants for complete case analysis with all predictors (2,966 for models without T1DGRS), well above the minimum required sample size.

Analysis was performed using Stata version 16.0 statistical software (StataCorp, College Station, TX) and R studio version 1.4.1106 (RStudio, PBC, Boston, MA). Models were developed using logistic regression on a complete case basis. We assessed the linearity of the relationship between continuous model variables and outcome on the logit scale using gamplot in R, with and without adjustment for covariates. As all relationships were linear with inclusion of covariates, model covariates were not transformed. GADA, IA-2A (and ZnT8A) were coded as the number of positive islet autoantibodies. Models were built in the following stages in the primary analysis: 1) clinical characteristics alone (age of diagnosis, BMI, sex, HDL, waist circumference), 2) clinical characteristics plus islet autoantibody number (GADA, IA2), 3) clinical characteristics plus islet autoantibody number plus T1DGRS, and 4) clinical characteristics plus T1DGRS.

To maximize utility for both clinical and research use, additional models were developed with and without waist circumference, HDL-C, and ZnT8 islet autoantibody (Supplementary Materials). Models were developed initially without race and ethnicity to allow for generalizability, but final models were tested in separate racial/ethnic subgroups to check for differences in performance. We assessed whether each of race and ethnicity, ZnT8 autoantibodies, and a T2DGRS improved model performance using the likelihood ratio test, Akaike information criterion, and Bayesian information criterion.

We assessed model discrimination using area under the receiver operating characteristic curve (AUCROC) analysis and precision recall curves and assessed model calibration visually with calibration plots using the Stata module PMCALPLOT (33). We also report model sensitivity, specificity, positive and negative predictive values, and overall classification accuracy for an illustrative binary model cutoff of 50% probability (≥50% assumed retained C-peptide, <50% assumed low C-peptide) and compared this with the accuracy of islet autoantibody status alone and latest provider type, with overall classification accuracy compared using the χ² test.

We then assessed model discrimination and calibration for all models separately by race and ethnicity (grouped into Black, White-Hispanic, and non-Hispanic White as above). To determine model performance in situations where clinical diagnosis is challenging and assess whether models had utility over and above clinical diagnosis and autoantibody testing, we tested models in the following participant subgroups: 1) older youth (onset age ≥10 years) with obesity (BMI z-score >1.64), 2) participants with negative and positive islet autoantibodies (GADA and IA-2A), 3) participants with a provider diagnosis of T1D and T2D, and 4) within the previously proposed four etiological subtypes of diabetes based on islet antibody status (GADA- or IA2A-positive or both negative) and presence/absence of insulin resistance (28,34).

Model internal validation was undertaken using bootstrapping (with replacement method, 500 bootstraps) using the Regression Modeling Strategies package for R version 6.3 (35).

Data from the SEARCH study is publicly available though the NIDDK central repository (https://repository.niddk.nih.gov/studies/search/), with additional data (eg, genetic risk scores) available through application to the SEARCH study steering committee.

A total of 2,991 participants in SEARCH had measured C-peptide between 3 and 10 years’ diabetes duration, of whom 2,966 met analysis inclusion criteria with available data for the routine clinical features model (model 1) and clinical characteristics with islet autoantibodies model (model 2). These features and T1DGRS were available for 1,918 participants (models 3 and 4). A flow diagram is shown in Supplementary Fig. 1. Overall characteristics of participants included in each model are shown in Supplementary Table 1. For participants included in models 1 and 2, median (interquartile range [IQR]) diabetes duration at C-peptide assessment was 74 (62–94) months; median (IQR) duration at assessment of model parameters (including islet autoantibodies) was 11 (5–18) months. For participants included in models 1 and 2, race and ethnicity was 65% non-Hispanic White, 14% White-Hispanic, 16% Black, and 4% other non-White. Characteristics of participants excluded from models because of missing T1DGRS are shown in Supplementary Table 2. Participants with missing T1DGRS were more likely to retain C-peptide (20 vs. 15%) and have other features of T2D.

Characteristics of participants included in models without T1DGRS, by C-peptide outcome status, are shown in Table 1. Retained C-peptide ≥250 pmol/L (≥0.75 ng/mL) was observed in 17% of participants. Those with retained C-peptide had characteristics largely consistent with T2D. T1DGRS and T2DGRS were similar in those with low and high retained C-peptide to values in those with diabetes type defined by a combination of latest provider classification and islet autoantibody testing (Table 1). Fasting C-peptide <250 pmol/L (<0.75 ng/mL) was highly specific for insulin treatment (specificity >99.9%). In contrast, 47.4% of those with C-peptide ≥250 pmol/L were treated with insulin.

Table 2 shows discrimination performance of models with clinical characteristics alone (age at diagnosis, BMI, waist circumference, sex, HDL-C) (model 1), clinical characteristics with GADA and IA-2A (model 2), and clinical characteristics and T1DGRS with and without islet autoantibodies (models 3 and 4, respectively). Performance of islet autoantibodies alone and provider diagnosis is also shown for comparison. Model coefficients are shown in Supplementary Table 3.

Models combining routine clinical features, with or without islet autoantibody testing (GADA, IA-2A) and T1DGRS, were highly discriminative for retained C-peptide (Table 2). AUCROC was 0.95 for clinical characteristics alone (model 1), rising to 0.97 with the inclusion of islet autoantibodies (model 2). AUCROC remained similar with the addition of T1DGRS (0.98, model 3). A model combining clinical features and T1DGRS (without autoantibodies) (model 4) also showed good discrimination (AUCROC 0.96). In those with T1DGRS available, model fit improved with addition of T1DGRS (Supplementary Table 4).

All models showed good discrimination (Fig. 1), and there were few participants with intermediate probabilities in models including islet autoantibodies or T1DGRS. Calibration was good across all models (Fig. 1). Precision recall curves for each model are shown in Supplementary Fig. 2; area under the precision recall curve ranged from 0.83 (model 1) to 0.92 (model 3).

All models had higher overall discrimination (AUCROC) than provider type or autoantibodies alone (Table 2 and Supplementary Table 5 [for participants with complete data for all models]). Using an arbitrary cutoff of 50% probability, all models retained substantially higher overall accuracy than islet autoantibodies alone (P < 0.0001 for all).

All models had similar performance in internal validation, with very low levels of optimism (optimism <0.003 for AUCROC and <0.033 for calibration slope), suggesting little error due to overfitting (Supplementary Tables 5–8).

Prevalence of retained C-peptide varied markedly by race and ethnicity: 45%, 28%, and 6% of those reporting Black, White-Hispanic, and non-White Hispanic race and ethnicity retained C-peptide ≥250 pmol/L. Discrimination and classification accuracy of models 1–4 in these ethnicities, compared with islet autoantibodies and clinical provider diagnosis alone, are shown in Supplementary Table 9. For models without T1DGRS, discrimination performance was lower in non-Hispanic White participants (models 1 and 2 AUCROC 0.88 and 0.94, respectively) in contrast to Black (AUCROC 0.95 and 0.97) and White-Hispanic (AUCROC 0.94 and 0.97) participants. Across all ethnicities, AUCROC and accuracy using a single (50% probability) cutoff was substantially higher for models incorporating islet autoantibodies than clinical diagnosis or autoantibody status alone.

Model calibration by race and ethnicity (Supplementary Fig. 3) showed that model 1 (routine clinical features) moderately underestimated probability of retained insulin secretion for Black participants with intermediate probability but had good calibration for most participants with very high or low probability. In contrast, in non-Hispanic White participants, probability of retained endogenous insulin secretion was moderately overestimated, consistent with the very low prevalence of retained C-peptide (and reported T2D) in this group. Calibration was good across all ethnicities in models incorporating islet antibodies and/or T1DGRS (Fig. 2) (model 2) and Supplementary Fig. 3 (models 1, 3, and 4).

Performance of models incorporating race and ethnicity are shown in Supplementary Table 10 and Supplementary Fig. 4. While incorporating race and ethnicity modestly improved model fit, adding them did not meaningfully change the overall discrimination.

In youth with age at diagnosis ≥10 years and obesity (BMI z-score >1.64) (a challenging group to classify), models retained high performance, with AUCROC 0.86, 0.95, 0.96, and 0.91 for models 1–4, respectively, and high calibration (Supplementary Table 10 and Supplementary Fig. 6). In this group (of whom 70% retained C-peptide), overall accuracy of provider type for retained C-peptide was modest (87%) and superseded by islet autoantibodies alone (accuracy 90.1%) and (using an illustrative 50% probability cutoff) models incorporating islet autoantibody testing (accuracy 92%) (P vs. provider diagnosis <0.05 for all).

In participants with negative autoantibodies (53% [452 of 853] of whom retained C-peptide), models 2 and 3 maintained high discrimination (AUCROC 0.95 [95% CI 0.94, 0.97] and 0.96 [0.95, 0.98], respectively) and good calibration (Supplementary Fig. 7). Within those with positive islet autoantibodies, models 2 and 3 also maintained good discrimination (AUCROC 0.86 [0.80, 0.92] and 0.87 [0.78, 0.950]) and reasonable calibration (Supplementary Fig. 7). However, only 49 of 2,213 included participants positive for islet autoantibodies retained C-peptide (3.2% and 1.4% of those with one and two positive islet autoantibodies, respectively).

Within those with clinically diagnosed T1D and T2D (latest available provider type), models retained the ability to discriminate those with and without retained C-peptide. In those diagnosed as having T1D (of whom 100 of 2,496 [model 1 and 2] and 57 of 1,640 [models 3 and 4] retained C-peptide), AUCROC was 0.84 (95% CI 0.78, 0.89), 0.90 (0.86, 0.93), 0.92 (0.88, 0.97), and 0.90 (0.86, 0.94) for models 1–4, respectively. In those diagnosed as having T2D (393 of 450 [models 1 and 2] and 228 of 263 [models 3 and 4] retained C-peptide), AUCROC was 0.81 (0.76, 0.87), 0.87 (0.82, 0.93), 0.88 (0.78, 0.95), and 0.82 (0.72, 0.91). Within those with four subtypes of diabetes defined by the SEARCH classification system (based on autoantibody status and insulin resistance) models also maintained high discrimination for retained C-peptide (Supplementary Table 11).

The addition of ZnT8 autoantibodies (available in 2,767 and 1,888 included participants for models 1 and 2 and 3 and 4, respectively) improved model fit but did not meaningfully change overall discrimination (Supplementary Table 12). The addition of a T2DGRS marginally improved model fit, with no impact on overall discrimination (Supplementary Table 13). Models without waist circumference and HDL maintained high performance (AUCROC 0.94–0.98) (Supplementary Table 14). Use of a 30-SNP T1DGRS in place of a 67-SNP score did not meaningfully change performance of models incorporating T1DGRS, with essentially identical model discrimination (Supplementary Table 15).

An online calculator (beta version) is available on https://julieanneknupp.shinyapps.io/SEARCH_ClassificationModel/. Model coefficients for all models are shown in Supplementary Table 16.

We have developed and evaluated prediction models to estimate the probability of a patient with youth-onset diabetes maintaining long-term C-peptide at levels associated T2D and its treatment requirements. These models show high performance, outperforming autoantibodies or clinical diagnosis alone, and could assist classification and treatment decisions in clinical practice. Importantly, models retained high performance in hard-to-classify groups, within those with T1D or T2D defined by clinical diagnosis or etiological markers, and within those of different race and ethnicity.

To our knowledge, this study is the first to develop a prediction model for retained C-peptide or diabetes classification in children and adolescents. Our findings are consistent with the high performance of models developed in adult-onset diabetes for classifying diabetes subtype based on retained C-peptide, or pancreatic histology, in a European/American population and on a combination of islet autoantibodies and C-peptide in a Chinese population (21–24,36). These previous models were developed on cross-sectional data, and the available data close to diagnosis in SEARCH allow prospective prediction of retained C-peptide, a key advantage, with the data available in SEARCH also allowing inclusion of a wider range of features.

The robust performance of models, including T1DGRS in those of different ethnicities, and improvement in calibration with this marker provide additional evidence for use of this score in multiple racial and ethnic groups, consistent with recent work demonstrating utility across ethnicities (37). However, gains seen with a T1DGRS are more modest than previously reported likely because the combination of clinical features and autoantibodies close to diagnosis provide very high accuracy, with limited room for further improvement in overall discrimination. As previously demonstrated, the utility of these scores is greatest in those who are hard to classify based on other features, for example, among those with islet autoantibody–negative suspected T1D or with single-positive antibodies and suspected T2D (38,39).

Strengths of this research include a large sample size and availability of prospective data. The large sample size available in SEARCH, combined with the modest number of predefined predictors included, means that the risk of overfitting (a common problem in prediction model development) is low, as evidenced by our internal validation. Additional strengths include diversity in race and ethnicity, allowing assessment of model performance across three racial and ethnic groups.

A limitation is the lack of other data sets for external validation. While robust internal validation suggests high performance without overfitting, further validation assessing model performance in different settings is an important area of further research. In particular, assessment in other racial/ethnic groups is needed. For example, our findings of high performance across different races and ethnicities may not apply to those of South Asian or Native American heritage, who have a particularly high risk of youth-onset T2D. Additionally, the T1DGRS used was developed in a White European population; thus, it is possible that ancestry-specific scores may further improve performance (40).

The outcome of these models was based on a C-peptide threshold associated with discriminating T2D from T1D and identifying lack of insulin requirement in previous insulin withdrawal studies. These thresholds are derived predominantly from adult studies, and there are limited data in children. A modestly higher threshold has been proposed for identifying T2D at diagnosis in a large Swedish study (41), but data for long-standing diabetes are lacking. The extremely high specificity of this threshold for insulin requirement in SEARCH gives confidence that children developing C-peptide below our cutoff will require insulin. However, it is possible that a higher threshold may be appropriate for identifying patients who benefit from treatment without insulin, as youth developing T2D generally have more obesity and insulin resistance than adults with this condition (42).

These models have the potential to assist with the diagnosis of T1D and T2D when patients first present and inform related decisions such as the need for insulin treatment and use of noninsulin glucose-lowering therapies. Models using only features available at diagnosis appear to perform as well as, or better than, latest clinical diagnosis in long-standing diabetes in specialist centers and could be implemented at no cost. Models may also help to identify uncertainty in classification and, therefore, target further assessment of classification biomarkers and the need for increased monitoring and safeguarding advice, for example, where a patient is initially suspected to have T2D and treated without insulin. When autoantibody or T1DGRS results are available, models allow interpretation of the result in the context of prior probability (based on other features), which is critically important to optimal use of these tests.

Importantly, tools of this nature are decision aids to supplement clinical judgment, which will incorporate additional information important to decision making, for example, glycemia and patient preference. The provision of continuous probabilities, rather than a binary result, allows clinicians to recognize uncertainty and act accordingly. Optimal use is likely to be through a staged approach at diabetes diagnosis, with classification biomarker testing prioritized in those with intermediate probability. In long-standing diabetes, C-peptide can be measured directly at modest cost, and these models may potentially assist in prioritization of testing. Models may also be useful for researchers who need to identify diabetes subtypes in research cohorts and may have particular utility in data sets where robust clinical diagnosis is not available and/or where genetic but not islet autoantibody data are available (24). In conclusion, prediction models combining routine clinical measures, with or without islet-autoantibodies and a T1DGRS, can accurately identify youth with diabetes who maintain endogenous insulin secretion in the range associated with T2D treatment requirements.

Note Added in Proof. Between initial online publication and final online and print publication, Seth A. Sharp was added as an author.

Acknowledgments. The SEARCH study investigators are indebted to the many youth and their families and health care providers whose participation made this study possible.

Funding. This study was supported by the Kaiser Permanente Southern California Marilyn Owsley Clinical Research Center (funded by Kaiser Foundation Health Plan and supported in part by the Southern California Permanente Medical Group), Clinical and Translational Research Institute at the Medical University of South Carolina (National Institutes of Health [NIH]/National Center for Advancing Translational Sciences [NCATS] grants UL1 TR000062 and UL1 TR001450), Seattle Children’s Hospital (grants U01 DP000244, U18 DP002710-01, U58/CCU019235-4), University of Washington (NIH/NCATS grant UL1 TR00423), University of Colorado Pediatric Clinical and Translational Research Center (NIH/NCATS grant UL1 TR000154), Barbara Davis Center at the University of Colorado at Denver (Diabetes and Endocrinology Research Center NIH grant P30 DK57516), University of Cincinnati (NIH/NCATS grants UL1 TR000077 and UL1 TR001425), and Children with Medical Handicaps program managed by the Ohio Department of Health. The SEARCH 4 study is funded by the NIH National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) (grants 1R01 DK127208-01 and 1UC4 DK108173) and supported by the Centers for Disease Control and Prevention. The Population Based Registry of Diabetes in Youth Study is funded by Centers for Disease Control and Prevention grant DP-15-002 and supported by NIH NIDDK grants 1U18 DP006131, U18 DP006133, U18 DP006134, U18 DP006136, U18 DP006138, and U18 DP006139. The SEARCH 1–3 studies are funded by Centers for Disease Control and Prevention grants 00097, DP-05-069, and DP-10-001 and supported by NIDDK. NIH funding supported the Kaiser Permanente Southern California (grants U48/CCU919219, U01 DP000246, and U18 DP002714), University of Colorado Denver (grants U48/CCU819241-3, U01 DP000247, and U18 DP000247-06A1), Cincinnati Children’s Hospital Medical Center (grants U48/CCU519239, U01 DP000248, and 1U18 DP002709), University of North Carolina at Chapel Hill (grants U48/CCU419249, U01 DP000254, and U18 DP002708), Seattle Children’s Hospital (grants U58/CCU019235-4, U01 DP000244, and U18 DP002710-01), and Wake Forest University School of Medicine (grants U48/CCU919219, U01 DP000250, and 200-2010-35171). R.A.O. is funded by a Diabetes UK Harry Keen Fellowship (16/0005529). J.K. is funded by Diabetes UK (21/0006328). A.G.J., B.M.S., and R.A.O. are supported by the National Institute for Health and Care Research (NIHR) Exeter Biomedical Research Centre. M.J.R.’s work on this analysis was supported by NIH NIDDK grant R01 DK124395.

The views expressed are those of the authors and not necessarily those of the NIHR or the Department of Health and Social Care. This study includes data provided by the Ohio Department of Health, which should not be considered an endorsement of this study or its conclusions.

Duality of Interest. R.A.O. is a co-investigator on a Randox R&D research grant. The study has received translational industry-academic funding from Randox R&D related to autoimmune genetic risk scores for prediction and classification of disease. There are no established patents, loyalties, or licensing agreements related to this grant. It is a 3-year grant (February 2022–2025). The approximate value is a £2.2 million program grant on genetic risk scores across autoimmune disease. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. A.G.J. wrote the first draft of the manuscript. A.G.J., B.M.S., R.A.O., J.D., and M.J.R. designed the study. A.G.J., B.M.S., J.K., and J.D. analyzed the data. R.A.O., D.M.D., W.A.H., S.A.S., E.L., A.S.S., A.K.M., R.B.D., A.W., S.M.M., C.P., and J.D. researched the underlying data. All authors provided helpful discussion and reviewed and edited the manuscript. A.G.J. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. Parts of this study were presented in abstract form at the 82nd Scientific Sessions of the American Diabetes Association, New Orleans, LA, 3–7 June 2022.