2214-P: Biomarker Validation Reveals Substantial Misclassification of Adult-Onset Type 1 Diabetes by EHR-Based Algorithms: Implications for Research Accuracy and Clinical Care

Introduction and Objective: Electronic health records (EHRs) underpin diabetes research and trial recruitment yet widely used EHR-based algorithms have not been robustly validated against gold-standard biomarker-defined diabetes subtypes. We aimed to provide the first large-scale biomarker validation of EHR-based diabetes classification algorithms for adult-onset T1D. Methods: We studied 7081 adult-onset diabetes cases (3365 insulin-treated and 3152 non-insulin-treated) from the MyCode cohort. Random C-peptide and islet autoantibodies (GADA, IA-2A, ZnT8A) were measured outside the honeymoon period at a median diabetes duration of 9.0 years (91% 3 year duration). We defined type 1 diabetes (T1D) as insulin treatment with low C-peptide (200 pmol/l) or autoantibody positivity. We defined type 2 diabetes (T2D) as negative islet antibodies with preserved C-peptide (600 pmol/l for insulin-treated individuals or absence of insulin treatment for at least 3 year). We evaluated eight widely used EHR-based classification algorithms for positive predictive value (PPV), negative predictive value (NPV), and ROC AUC. Results: Mean age at diagnosis was 59.8 years (SD 14.9); 55% were female, and 95% were Europeans. Biomarkers identified 3.1% adult-onset T1D and 71% T2D. All algorithms showed poor T1D identification, with PPV ranging from 23-79% despite high NPV (96-97%). The lipid model by Lynam et al achieved the highest PPV at 79%. Discrimination was modest across all algorithms (ROC AUC 0.65-0.78), including 0.66 for the widely used eMERGE algorithm, and performance was worse in insulin-treated individuals for all algorithms (ROC AUC reduced to 0.62-0.73, NPV 96-97% and PPV 49-77%). Conclusion: EHR-based algorithms substantially misclassify biomarker-defined adult-onset T1D. These findings raise caution against use of current EHR tools for studying adult T1D and strongly support the need for biomarker-informed classification strategies. Disclosure J. Li: None. R. Stahl: None. A. Golden: None. J.S. Haley: None. A. Jones: None. R.A. Oram: Other - Research support, The University of Exeter has a licensing and royalty agreement for a 10 SNP T1D GRS with Randox; Current; Randox. Other - Advisory Panel, Consulting, Invited talks; Current; Sanofi. Other - invited talk, internal teaching talk on genetics; Ended; Novo Nordisk. D. Carey: None. K. Patel: None. U. Mirshahi: None. Funding Breakthrough T1D (3-SRA-2022-1241-S-B)