 Congenital heart defects affect approximately 1% of all babies born each year and account for almost 20% of all newborn deaths. Early diagnosis while still in the womb can greatly improve an affected baby's chance of survival. Unfortunately, diagnosis relies exclusively on ultrasound imaging, where accurate readings aren't guaranteed. Researchers in Japan are tackling this problem by enlisting the help of artificial intelligence. More importantly, they're helping the doctors entrusted with patient care to understand how AI programs spot heart defects. Advancements in artificial intelligence have improved how congenital heart defects are diagnosed. Ultrasound videos of fetal hearts beating normally and others with structural defects can be studied with AI, which can then determine whether the fetal hearts in new videos are abnormal or not. However, many medical professionals are wary of adopting this approach because of the so-called black box problem. The rules by which AI reaches its conclusion are so complex that human users, including the developers themselves, are often unable to understand the rationale. Researchers at the Rekin Center for Advanced Intelligence Project and colleagues set out to demystify the AI decision-making process. Their goal was to create a visual representation of the decisions that their AI made, which could then be used to support the ultrasound screening process in the clinic. This graph chart diagram is created in two steps. First, various structures in the cardiovascular system are detected through ultrasound anatomy scans, including abdomen, 4-chamber view, or 4CV, and 3-vessel trachea view, or 3VTV. Then, these data points are plotted on a two-dimensional plane. The shape of the diagram is compared to that obtained for a normal heart in vessels, and this information is used to generate an abnormality score. To test this approach, the team compared the diagnostic accuracy of medical experts, fellows, and residents when using ultrasound videos alone, and when used in combination with graph chart diagrams and abnormality scores, but with the actual decision of the AI withheld. They found that correct diagnoses increased across all levels of medical experience when the AI-generated data were used. The greatest improvement was seen among residents, who became 13% more accurate in their diagnoses. However, they were still 12% less accurate than the AI alone, emphasizing the importance of experience even when AI applications are available. Of the three groups, fellows who most often perform fetal cardiac screening in hospitals seem to benefit the most from the AI data. Although further studies that incorporate other types of ultrasound equipment and data are needed, these results highlight how AI technology can benefit patients and clinicians and could help facilitate adoption by medical professionals by shedding light on the black box.