Table 3

Summary of repeated train-test split results for selection of the most significant predictors of C-peptide detection

Probability of inclusionCoefficient from modelORP
Onset age1.00β11.110<0.0001
Duration1.00β20.701<0.001
GRS0.80β30.7480.061
Duration * IA–2A0.72β40.8960.175
GRS * ZnT8A0.65β51.3800.057
GADA0.58β61.2300.128
Onset age * ZnT8A0.52β71.080<0.001
Duration * ZnT8A0.51β80.9080.190
IA–2A0.47β91.6800.041
ZnT8A0.39β100.6250.198
  • Disease duration; onset age; GRS; titers for IA–2A, ZnT8A, and GADA; and all two-way interaction effects of predictor variables were tested for inclusion in the C-peptide model simultaneously with penalized logistic regression with repeated 10-fold cross validation for feature selection. The data set was repeatedly partitioned in half for training and testing, and the probability of each predictor (including all two-way interaction effects) being in the model was calculated over 1,000 iterations. All coefficients with at least 50% probability of inclusion and their main effects are reported. Interaction effects are denoted as two variables separated by an asterisk. The glmnet, version 3.0, package in R was used for penalized logistic regression and feature selection.

  • Regression coefficients (βi) are reported in exponentiated form as ORs, such that OR = eβ. The coefficient values correspond to the β values from the overall model formula.