Summary
A single-centre retrospective study found routine clinical data could help rank likely genetic conditions in children with congenital heart disease. The AI models were evaluated in a selected cohort, so the results are a research finding rather than evidence of clinical diagnostic performance.
A retrospective study of 1,720 children with congenital heart disease found that artificial-intelligence models using routine clinical data could rank likely genetic conditions among their top three suggestions. Using information available at the echocardiogram visit, class-balanced logistic regression reached a balanced top-three accuracy of 0.758; adding later medical-record data, radiology notes and ICD-10 codes raised the score to 0.809. The findings are reported in a medRxiv preprint.
The researchers compared logistic regression with a multimodal deep-learning network, which combined different kinds of inputs, including echocardiogram images and clinical records. The study examined whether information already collected during care could help clinicians identify which genetic conditions to investigate.
How the models performed
The cohort included 988 children with common congenital-heart-disease-associated genetic conditions, 231 with rare genetic diagnoses and 501 whose cytogenetic testing and congenital-heart-disease gene sequencing were negative. The researchers added echocardiogram images and serial medical-record data to the model in the order those inputs would become available in clinical care.
The primary measure, balanced top-three accuracy, counted a diagnosis as correctly ranked if it appeared among the model’s three leading suggestions, then averaged performance across conditions. This gives each condition a role in the score rather than allowing the most frequent conditions alone to determine it. With inputs from the echocardiogram visit, logistic regression scored 0.758, compared with a reported chance value of 0.429. With later records, radiology notes and ICD-10 codes added, the score rose to 0.809. The multimodal network performed similarly at each stage and did not improve accuracy over logistic regression.
The team also tested whether the network’s learned representation could help with a condition excluded from training. In that analysis, children with the withheld condition were placed nearest to others sharing a gene group in 0.37 of cases, compared with 0.10 without the embedding. The reported difference was +0.268, with a 95% confidence interval of +0.242 to +0.295. This suggests the representation may help find similarities to known genetic groups even when a particular condition was not included in training.
A research tool for prioritising diagnoses
Genetic conditions can affect care for children with congenital heart disease, while some diagnoses are missed or made late. A system that uses information available early in care could help prioritise which conditions warrant closer evaluation. The study’s findings support further investigation of that approach, including the potential of similarity-based retrieval for uncommon conditions.
The results come from a retrospective, single-centre cohort selected for either a confirmed genetic diagnosis or negative results on specified genetic testing. They measure how well the models ranked conditions in that study population, not how accurately they would diagnose children prospectively in routine practice. The model is therefore best understood as a research method for generating a shortlist, not as a genetic diagnosis or a replacement for clinical evaluation.