Type 1 diabetes prediction in autoantibody-positive individuals: performance, time and money matter.

AIMS/HYPOTHESIS: Efficient prediction of clinical type 1 diabetes is important for risk stratification and monitoring of autoantibody-positive individuals. In this study, we compared type 1 diabetes predictive models for predictive performance, cost and participant time needed for testing.

METHODS

We developed 1943 predictive models using a Cox model based on a type 1 diabetes genetic risk score (GRS2), autoantibody count and types, BMI, age, self-reported gender and OGTT-derived glucose and C-peptide measures. We trained and validated the models using halves of a dataset comprising autoantibody-positive first-degree relatives of individuals with type 1 diabetes (n=3967, 49% female, 14.9 ± 12.1 years of age) from the TrialNet Pathway to Prevention study. The median duration of follow-up was 4.7 years (IQR 2.0-8.1), and 1311 participants developed clinical type 1 diabetes. Models were compared for predictive performances, estimated cost and participant time.

RESULTS

Models that included metabolic measures had best performance, with most exhibiting small performance differences (less than 3% and p>0.05). However, the cost and participant time associated with measuring metabolic variables ranged between US$56 and US$293 and 10-165 min, respectively. The predictive model performance had temporal variability, with the highest GRS2 influence and discriminative power being exhibited in the earliest preclinical stages. OGTT-derived metabolic measures had a similar performance to HbA- or Index-derived models, with an important difference in cost and participant time.

CONCLUSIONS/INTERPRETATION: Cost-performance model analyses identified trade-offs between cost and performance models, and identified cost-minimising options to tailor risk-screening strategies.