1690-P: Predicting Onset of Type 1 Diabetes Using a Novel Interaction Model

Abstract

Progression to type 1 diabetes (T1D) is strongly predicted by multiple risk factors including genetic and non-genetic variables such as islet autoantibodies (AAb). Little is known, however, about the impact of interactions among the known risk factors on T1D risk prediction. Through consideration of interactions among risk factors, we aimed to evaluate time-to-diagnosis for T1D from seroconversion for AAb using a novel time-to-event model (RankSvx). RankSvx simultaneously predicts subject-level event times and risk scores as measured by Mean Absolute Error (MAE) and Concordance Index (CI), respectively. We control for subject-level characteristics (4 HLA groups, sex, and their interactions with AAb, age at seroconversion and visit-level AAb count) and use LASSO for feature selection. We also explored up to 12 months of AAb dynamics (i.e., appearance or disappearance of AAb in 3-month intervals) post seroconversion using an interaction model and compared it to a non-interaction (baseline model at seroconversion) and a traditional Cox model. We used a harmonized dataset of the DAISY, DiPiS, and DIPP studies with 1,880 AAb positive subjects of whom 512 (27%) progressed to T1D. First, we found the interaction model was superior to the baseline model without interactions (Wilcoxon Rank test p-value: 0.016). Second, incorporating AAb dynamics further improved T1D onset time prediction. When including post seroconversion AAb dynamics at 12 months, the interaction model performed 15% better than the baseline model (CI: 0.83 vs. 0.68, MAE: 2.5 years). Finally, the top identified interactions included: 1) young age at seroconversion interaction with IA-2A, and 2) HLA Group that includes at least one DR3-DQ2.5 haplotype interaction with ZnT8A or GADA. Our results are consistent with recent findings in the TEDDY dataset of patterns of AAb appearance and genetic association and age-related T1D incidence. Further investigations on seroconversion and post seroconversion endotypes are necessary. Disclosure B. Liu: None. V. Anand: None. M. Ghalwash: None. K. Ng: None. J.L. Dunne: None. R. Veijola: None. M. Rewers: None. M. Lundgren: None. Funding JDRF