Taksøe-Vester CA, Mikolaj K, Petersen OBB, Vejlstrup NG, Christensen AN, Feragen A, Nielsen M, Svendsen MBS, Tolsgaard MG. Ultrasound Obstet Gynecol. 2024 Jul;64(1):36-43. doi: 10.1002/uog.27608. Epub 2024 Jun 3.PMID: 38339776
Take home points:
- The authors developed an artificial intelligence (AI) model capable of identifying fetuses who are at risk of developing coarctation of aorta (CoA).
- The MPA/Aao ratio emerged as the most crucial individual feature, setting the threshold at 1.15 in MPA/Aao ratio resulted in a sensitivity of 90.3% and a specificity of 61.9%.
Commentary from Dr. Vimal Jayswal (Indiana, USA), editor of Pediatric & Fetal Cardiology Journal Watch.
Introduction
Coarctation of the aorta (CoA) is the most common under diagnosed congenital heart defect during prenatal screening, with a true positive detection rate of <50%. High false positive and false negative rates seriously affect prenatal consultation and postnatal management, leading to unnecessary parental anxiety and hospitalization costs and potential delays in critical care. This study demonstrated the potential of artificial intelligence (AI) models in enhancing the prenatal screening, particularly in resource constraint settings. As AI is taking over every field, including medicine, the authors suggest using this AI model as an initial screening tool at 18-22 weeks of gestation, to improve detection rates of CoA.
Coarctation of aorta screening: The main limitations to the prenatal diagnosis of CoA are the low specificity of the classic echocardiographic signs (ventricular disproportion, great vessel asymmetry, shelf of the aortic isthmus), individual variation, image quality and limited access to MFM experts. The authors developed an AI algorithm to recognize cardiac standard planes and perform automatic biometric measurements. The hypothesis was that performing automated biometric measurements during screening examinations would lead to more accurate identification of fetuses at risk of developing CoA.
Data collection and its clinical relevance
This study was a national retrospective observational study across multiple centers, Danish prenatal ultrasound screening program between 1st January 2008 and 31st December 2018. All cases that received a postnatal diagnosis of CoA during the study period were included and matched them with healthy controls at a 1:100 ratio. This study used cardiac biometrics obtained from the four-chamber and three-vessel views in a logistic regression-based prediction model. The right ventricle (RV) area and length; left ventricle (LV) diameter and the ratios of RV/LV areas and main pulmonary artery/ascending aorta diameters were measured automatically via AI algorithms in eight fetal cardiac standard planes. Using logistic regression and backward feature selection, the study prediction model had an area under the ROC curve of 0.96 and a specificity of 88.9% at a sensitivity of 90.4%.
Results
99 fetuses with a postnatal diagnosis of CoA born between 2008 and 2018 were identified. Of these, 26 were excluded from further analysis.
Fetuses that had been diagnosed with CoA postnatally displayed significant deviations from healthy controls in terms of their cardiac structure. Specifically, these fetuses had smaller left cardiac structures, such as the mitral valve diameter and atrial and ventricular dimensions, while also exhibiting larger right cardiac structures in terms of right ventricular (RV) dimensions (Table 2). Moreover, the CoA fetuses showed a significantly larger diameter of the MPA, a smaller diameter of the Aao, and larger ratios of RV/left ventricular (LV) area and MPA/Aao diameter when compared with controls (Table 2).
The MPA/Aao ratio emerged as the most crucial individual feature, based on the AUC (0.90). Setting the threshold at 1.15 in MPA/Aao ratio resulted in a sensitivity of 90.3% (95% CI, 83.4–97.1%) and a specificity of 61.9% (95% CI, 60.8–63.1%)
Discussion
Utilization of AI models for prenatal detection of COA as an initial tool in the center with limited MFM expertise to obtain additional and challenging fetal echocardiogram images. AI models can standardize the prenatal screening process, identifying potential CoA and flagging cases for further monitoring. Furthermore, this early detection of COA AI predictive model could pave a way forward in creating a Telehealth fetal team to increase access to specialized care and democratizing the AI driven learning and screening tool for all. This multi-faceted approach would enable more accurate stratification and diagnosis of CoA, ultimately improving prenatal care and outcome.
Conclusion
In the current era, when AI is getting involved in every field, and changing the way we approach; this pioneer study is starting to bridge the gap between AI technology and fetal care for challenging diagnosis of prenatal diagnosis of COA.
By integrating a dedicated advance prenatal CoA screening protocol with Telehealth service, led by an MFM expert at a tertiary care center could enhance the next dimension in early prenatal detection of Coarctation of Aorta and its postnatal management.
