We used RF modeling to collectively assess seven T1D-implicated serological markers, in combination with AAbs, GRS2, and demographic information, for their utility in classifying study participants according to AAb/T1D status in a cross-sectional cohort. We observed higher GRS2 values, younger age, and reduced lipase levels associated with ≥2AAb+ versus AAb− classification. While speculative, these criteria may help identify AAb− individuals with the highest likelihood of progressing to ≥2AAb+ status, which could facilitate earlier implementation of low-risk clinical interventions. Further, reduced trypsinogen levels emerged as informative for differentiating recent-onset T1D from ≥2AAb+, altogether confirming our previous reports of reduced lipase and trypsinogen levels in ≥2AAb+ and T1D compared with AAb− individuals.1 26 31 GRS2 emerged as one of the most useful variables for differentiating AAb− and ≥2AAb+ study participants but was less informative when comparing the 1AAb+ versus ≥2AAb+ or ≥2AAb+ versus T1D groups.

While GRS2 is a stable metric ideal for T1D risk assessment early in the disease process, additional dynamic risk markers are needed to monitor the risk of disease progression. AAb specificities, especially ZnT8A and IAA, were the most informative for differentiating ≥2AAb+ versus 1AAb+ subjects. Accordingly, the presence of ZnT8A is considered indicative of more advanced insulitis,32 suggesting that ZnT8A+ individuals may benefit from more aggressive or tailored therapies. In contrast, AAb specificities were not informative for differentiating ≥2AAb+ individuals from those with recent-onset T1D. Yet, previous studies that showed IA-2A levels33–35 or emergence as a second AAb36 were predictive of progression from ≥2AAb+ to clinical onset. Study design may account for this discordance, as these longitudinal studies were conducted on younger, higher risk children. On the other hand, it is certainly possible that modeling AAb levels (vs status) or accounting for the order of AAb emergence would improve the ability of AAbs to differentiate ≥2AAb+ individuals from those with T1D in our models. Hence, the incorporation of AAb status rather than titers is a limitation of this study. In our cross-sectional assessment, RF4 had low accuracy, indicating difficulty in differentiating ≥2AAb+ from T1D. This is consistent with the well-established disease staging model: once an individual becomes positive for any two AAbs, they have a near 100% lifetime risk of progressing to stage 3 disease.37

We previously reported that associations between small pancreas size, low exocrine enzyme levels, and T1D progression hold true, even after adjustment for age.1 However, due to the effects of age on pancreas organ size and exocrine enzyme levels,26 27 we anticipated age to emerge as a significant predictor in classification and refrained from overinterpreting its impact on RF model outcomes. Nonetheless, age is well recognized as an important feature for T1D risk assessment, histopathological phenotype, and clinical subtype, as previously reviewed.38 39

In our models, certain variables with high RF importance scores were not significantly informative for logistic regression. This apparent discrepancy is actually the basis for our approach: unlike logistic regression, which assumes a linear relationship of independent variables to log odds, the RF algorithm can capture complex dependency patterns between the independent and dependent variables. Further, logistic regression depends on the assumption that little or no collinearity exists among the predictor variables. Unconstrained by these assumptions, the RF approach is less sensitive to the effect of outliers, more resistant to overfitting, and an overall more robust approach for our purposes.10 As such, discrepancies in variable significance between the RF and logistic regression models may be explained by assumptive constraints inherent to logistic regression rather than poor variable predictive capacity. Nonetheless, significantly informative variables using both modeling approaches represent high-interest markers for follow-up in longitudinal studies.

T1D prediction and risk assessment outside of White/European ancestry cohorts has, historically and unfortunately, been understudied. While self-reported race and ethnicity were not informative for T1D status classification in our models, we were insufficiently powered for robust subgroup analyses. Nevertheless, outreach efforts to enrich our biobank for individuals with racially diverse backgrounds are ongoing.40 Prandial status and sex were also among the least informative variables for T1D status classification across all four RF models. Certain variables did not meet the significance threshold in the RF models, such as IGF1 percentile, despite previous reports of association with T1D progression.3 41 One reason for this discrepancy could be that the change in IGF1 levels over time is more informative than cross-sectional measurements. Indeed, Shapiro et al demonstrated that IGF1 levels decreased over time in ≥2AAb+ individuals and those who progressed to clinical diagnosis but remained stable in 1AAb+ individuals.3 Hence, our RF models, which are limited by the cross-sectional nature of the cohort, may underestimate the predictive capacity of dynamic markers such as IGF1.

Other T1D risk models have been described. The Diabetes Prevention Trial–Type 1 Risk Score (DPTRS), which incorporated fasting C-peptide and glucose levels, oral glucose tolerance test (OGTT)-derived C-peptide measures, BMI, and age, identified AAb+ individuals with high likelihood of progression to clinical diagnosis within 2 years.42 Index60, which considers fasting C-peptide and OGTT-derived C-peptide and glucose measures, also identified AAb+ individuals at very high risk for clinical diagnosis,43 even surpassing HbA1c in predictive capacity.44 These models are beneficial for facilitating early detection of dysglycemia in at-risk (ie, AAb+) individuals in a clinical setting,43 and may even be used as a prediagnostic endpoint for clinical trials.45 Approaches that can be implemented earlier in the disease process are needed to identify: (a) AAb− individuals at greatest risk for seroconversion, and (b) 1AAb+ individuals most likely to progress to ≥2AAb+ status. The Combined Risk Score (CRS), which incorporated AAb (specifically IAA, GADA, and IA-2A), GRS2, and family history, modeled risk for progression to T1D clinical onset within the first 10 years of life, with the most pronounced utility in predicting seroconversion.46 Indeed, models like CRS and our RF models that account for the genetic burden of T1D in addition to AAb have the potential for earlier implementation compared with DPTRS and Index60, which are ideal for stratifying risk closer to clinical onset. Of note, the CRS did not consider ZnT8A, which we found to be the single most important predictor when differentiating 1AAb+ from ≥2AAb+ subjects. Further, the CRS, DPTRS, and Index60 models were generated from cohorts comprising entirely at-risk individuals and, therefore, may not fully account for the full heterogeneity of T1D.47

We tested whether a combination of serological, genetic, and demographic information could aid in the classification of AAb/T1D status in individuals ≤30 years old using four RF models in a cross-sectional cohort. We observed that markers above the significance threshold were differentially informative across the spectrum of disease progression: higher GRS2 and reduced lipase levels differentiated ≥2AAb+ (vs AAb−) study participants; the presence of ZnT8A, GADA, and IAA favored ≥2AAb+ (vs 1AAb+) classification; and reduced trypsinogen and elevated lipase levels were associated with increased likelihood of T1D (vs ≥2AAb+) classification. Future studies will ascertain the generalizability of these models and the positive predictive value of these markers, particularly lipase and trypsinogen in combination with GRS2 and AAb, in an independent longitudinal cohort.