Artificial Intelligence-Based Modeling to Aid in the Identification of Genetic Conditions in Infants with Congenital Heart Disease

Logistic regression and multimodal fusion (echocardiography, medical records) to shortlist likely genetic diagnoses

This retrospective single-center study included 1,720 children with congenital heart disease and an echocardiogram: 988 with a common CHD-associated genetic condition, 231 with a rare diagnosis and 501 with negative testing (cytogenetics and CHD gene sequencing). Logistic regression and a multimodal deep learning fusion network were fitted to the same inputs — echocardiogram images, then serial medical record data — added in the order they become clinically available. Using only inputs from the echocardiogram visit, logistic regression ranked the correct diagnosis among the top three with a balanced accuracy of 0.758 (chance: 0.429), rising to 0.809 with serial records, radiology notes and ICD-10 codes. Multimodal fusion performed similarly at each stage, but, for a condition withheld from training, its embedding placed affected children nearest to others sharing a gene group in 37% of cases, versus 10% without the embedding.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

Preprint on medRxiv, not peer reviewed. The clearest result goes against deep learning: plain logistic regression performs as well as multimodal fusion, and most of the information is available at the first echocardiogram. Retrospective single-center study without external validation; the tool points towards a likely genetic diagnosis but does not replace the exome or genome that will confirm it.

Analysis by Dr Thibaut Benquey