Journal: European Heart Journal. Digital Health

Article Title: Multimodal data integration to predict atrial fibrillation

doi: 10.1093/ehjdh/ztae081

Figure Lengend Snippet: Model development. ( A ) Numbers of participants from different data sources. Diagrams represent the numbers of participants from each of the four data sources for incident AF after Visit 3 (left) and prevalent AF at Visit 5 (right). The combined datasets are intersections among participants in the four data sources, consisting of the test set (consistent across the combined data and four data sources) and the combined training set. For all four data sources, we excluded non-White or Black participants, Black participants in Washington County and Minneapolis centres due to low numbers. We also excluded participants with atrial fibrillation rhythm on the source ECG or poor ECG quality. CRS, clinical risk score; PRS, polygenic risk score; ECG, electrocardiogram model; Prot, protein score. ( B ) Model development diagram. CRS represents the predicted score generated by logistic regression utilizing 11 clinical variables: age, race, height, weight, systolic blood pressure, diastolic blood pressure, current smoking, antihypertensive medication use, diabetes, prevalent heart failure, and prevalent myocardial infarction. PRS was derived with weights provided in a study based on AF GWAS across six cohorts. Combined models integrated information from different combinations of data source. For genotype, ECG, and proteomics, combined models incorporate the risk scores from each data source. For clinical variables, the combined models directly utilized the 11 clinical variables themselves instead of the CRS, proved to produce better prediction performance. ( C ) Diagram of data integration design. Test data remained consistent for comparing both separate and combined models within the same replicate. Sample sizes for incident AF after Visit 3 and prevalent AF at Visit 5 are listed for each dataset. In the case of separate training data for ECG, we additionally partitioned it to network training and validation sets by a ratio of 0.9:0.1. For separate training data regarding proteomics, we employed five-fold cross-validation to choose from 10 values of lasso penalty parameters.

Article Snippet: Participant plasma samples from Visit 3 and Visit 5 were stored at −80°C after initial collection and thawed for analysis with the Somalogic 5K aptamer-based proteomics platform in a central laboratory as previously described.

Techniques: Generated, Derivative Assay