1280-P: Improving Type 1 Diabetes (T1D) Prediction by Incorporating Growth Features into Landmark Models

We explored the association between growth features and T1D development using landmark analysis at different ages. Analysis included individuals from 2 birth cohort studies: DAISY and BABYDIAB (n=2,664; 129 progressed to T1D). Using height and weight measured over time, percentiles for age were calculated. Missing values were imputed using LMS parameters of CDC growth charts. Rates of change of percentiles were computed over the prior year. Twelve ages (1-12 years) were used for landmark analysis with Random Survival Forest to predict probability of T1D onset in the 1- to 19-year follow up windows. The baseline model used HLA risk group, sex, T1D family history and breastfeeding history. The full model added growth features: height and weight percentiles at age and change in percentiles. Performance was measured using C-index and feature importance was ranked. Incorporating growth features significantly improved prediction accuracy of T1D onset for 95% of combinations of landmark ages and prediction window sizes (Table 1, not all ages and windows shown). The order of features from most to least predictive is: HLA group, rates of height and weight changes, height and weight percentiles, family history, breastfeeding and sex. This analysis demonstrates that using growth features can significantly improve prediction of T1D. Table 1: C-Index with and without inclusion of growth features for landmark age of 3 years. Disclosure Z. Li: None. V. Anand: None. J.L. Dunne: None. B.I. Frohnert: None. W. Hagopian: Consultant; Self; Novo Nordisk Inc. H. Hyoty: None. M. Maziarz: None. A. Ziegler: None. J. Toppari: None. Funding JDRF (1-IND-2019-717-I-X, 1-SRA-2019-722-I-X, 1-SRA-2019-720-I-X, 1-SRA-2019-721-I-X, 1-SRA-2019-719-I-X, 1-SRA-2019-723-I-X)