Multimodal AI for Precise Clinical and Genetic Diagnosis of Congenital Heart Disease

Abstract: Congenital Heart Disease (CHD) is the most common and life-threatening birth defect, caused by pathogenic variants in over 400 genes, with hundreds of thousands of Americans affected. Despite up to 90% of cases thought to be due to genetic variants, currently only <30% cases are resolved. Genetic diagnosis is critical, as it modifies prognosis, management, medical and surgical therapy, and future family planning. A large array of data relevant for genetic diagnosis is typically available due to near universal use of DICOM image formatting for echocardiography, standardized classification language for CHD, wide electronic medical record (EMR) adoption, and whole genome sequencing. However, only a fraction of these data are currently being used effectively. The search for pathogenic variants is currently manual, making it hard to harness valuable information from previously resolved cases. Interpretation of the genome is limited to only the ~1% that codes for protein. To address these gaps, we will develop a multimodal AI module that harnesses information from related cases to prioritize genes likely to harbor pathogenic variants and a variant identification module to identify likely causative variants from whole genome sequencing data. The gene prioritization and variant identification modules will be co-designed with input from stakeholders, integrated into an automated process, and validated while addressing key ethical considerations. By completing this project, we will harness previously untapped information to automate genetic diagnosis in CHD, thus enhancing patient outcomes, advancing understanding of CHD, and paving the path toward wider clinical and genetic applications of multimodal AI.