1462-P: Genetic Risk Score for Type 1 Diabetes across Ethnic Populations via Multitask Learning

Introduction: Genetic risk scores (GRS) for the prognosis of Stage 3 type 1 diabetes (T1D) are potentially biased due to the predominant race/ethnicity (European) from which they were derived. We examined the performance of the current GRS2 model in The Environmental Determinants of Diabetes in the Young (TEDDY) study across different ethnic groups and proposed a fair GRS model using multi-task learning (MTL).

Methods: Through the lens of Multi-Task Learning (MTL), we incorporated different group lasso (GL) regularization strategies and have extracted commonly shared and different group-wise features among ethnic groups based on the identified single nucleotide polymorphisms (SNPs) set utilized by the published GRS2 model. Receiver operating characteristic area under the curve (ROC-AUC) is used to test performance by ethnicity.

Results: We studied 4,918 Europeans (ethnic majority), and 2,468 non-Europeans (ethnic minority) children; 274 ethnic majority and 116 ethnic minority children developed T1D. Our MTL-based strategy resulted in an improved GRS2 for each ethnic minority group. The ROC-AUC performance for the ethnic minority population demonstrated an improvement, increasing from 0.679 to 0.713, indicating a fairer prognosis performance compared to the 0.728 of the GRS2 for the ethnic majority population, which uses the 67 SNPs as features for different ethnic groups.

Conclusion: MTL with different regularization strategies for developing fair GRS models improved T1D prognosis performance for ethnic minority groups in TEDDY. Future efforts should aim to derive fair GRS models by considering fairness metrics when searching for predictive genetic markers.