Panel A shows a receiver operating characteristic (ROC) curve with sensitivity on the y-axis and 1-specificity on the x-axis. The curve represents the prediction of seropositive CeD using logistic regression. Panel B displays another ROC curve comparing three scoring methods: PRS alone, PRS plus family history, and PRS plus family history plus clinical diagnoses. Panel C features violin plots showing the distribution of AoU-CeD scores in seropositive CeD cases versus matched non-CeD controls, stratified by ancestry. Panel D presents violin plots validating the AoU-CeD score in CeD patients with variable tTG-IgA values. Each panel includes specific labels, legends, and annotations relevant to the data being presented.
The AoU-CeD score combines clinical risk factors and PRS to improve CeD prediction over and above that achieved with the PRS alone. (A) Logistic regression and machine-learning approaches were used to identify the variables that best predicted CeD. The input variables included sex, age, genetic ancestry, HLA-DQ genotype, and 42 clinical features, including symptoms, comorbid conditions, and complications related to malabsorption. Seropositive CeD participants were used as positive cases. A ROC curve was plotted, and the model coefficients for the top predictors are reported. (B) We evaluated the extent to which the incorporation of clinical data improved prediction over that achieved with the PRS alone, by testing models including family history (FH) and seven clinical diagnoses (Dx). ROC curves comparing three scoring methods: PRS alone, PRS + FH, and PRS + FH + Dx. Scores were calculated with a log-additive model, combining log-transformed values weighted by logistic regression coefficients. (C) Violin plots showing AoU-CeD score distributions in seropositive CeD cases vs. matched non-CeD controls, using a threshold capturing 90% of seropositive CeD patients. The AoU-CeD score was then applied to CeD patients without high levels of serological markers (>20 IU/ml), stratified by ancestry. (D) Validation of the AoU-CeD score in CeD patients from All of Us with available but variable tTG-IgA values recorded in their EHRs. The dashed line indicates the optimal cutoff for discriminating between cases and controls.